[NPU] Improve the performance of the embedding for NPU

Currently, the embedding operator is supported on the NPU and has passed the accuracy test. However, the running performance is much lower than that of other systems such as Hugging Face, and further improvements are needed.

```
********** Benchmark Data **********
[
{
"kernel_name": "embedding",
"kernel_provider": "liger",
"metric_name": "speed",
"metric_unit": "ms",
"gpu_name": "Ascend910B4",
"x_name": "V",
"x_label": "embedding dimension",
"x_values": [
1024,
2048,
4096,
8192,
16384,
32768,
65536,
131072
],
"y_values_50": [
43.11292266845703,
42.90717315673828,
43.2876091003418,
43.323081970214844,
42.62879943847656,
43.2666015625,
43.26323699951172,
43.450679779052734
],
"y_values_20": [
43.11175537109375,
42.90643310546875,
43.28743362426758,
43.32281494140625,
42.62324523925781,
43.2645263671875,
43.26128387451172,
43.43910217285156
],
"y_values_80": [
43.11408615112305,
42.90790939331055,
43.287784576416016,
43.32334518432617,
42.63435745239258,
43.2686767578125,
43.265193939208984,
43.462257385253906
],
"timestamp": "2026-01-21 03:48:04",
"kernel_operation_mode": "forward",
"extra_benchmark_config_str": "{\"B\": 32, \"T\": 512, \"D\": 768, \"dtype\": \"torch.float32\"}",
"liger_version": "0.0.0"
},
{
"kernel_name": "embedding",
"kernel_provider": "huggingface",
"metric_name": "speed",
"metric_unit": "ms",
"gpu_name": "Ascend910B4",
"x_name": "V",
"x_label": "embedding dimension",
"x_values": [
1024,
2048,
4096,
8192,
16384,
32768,
65536,
131072
],
"y_values_50": [
0.08064000308513641,
0.09166000038385391,
0.11357000470161438,
0.1482200026512146,
0.18525999784469604,
0.21186000108718872,
0.2272000014781952,
0.23625999689102173
],
"y_values_20": [
0.08020000159740448,
0.09121999889612198,
0.1130559965968132,
0.14762000739574432,
0.18479999899864197,
0.21121999621391296,
0.22665999829769135,
0.2358199954032898
],
"y_values_80": [
0.08157600462436676,
0.09262000024318695,
0.11438000202178955,
0.14905999600887299,
0.18602000176906586,
0.21264000236988068,
0.22774000465869904,
0.2370000034570694
],
"timestamp": "2026-01-21 03:48:17",
"kernel_operation_mode": "forward",
"extra_benchmark_config_str": "{\"B\": 32, \"T\": 512, \"D\": 768, \"dtype\": \"torch.float32\"}",
"liger_version": "0.0.0"
},
{
"kernel_name": "embedding",
"kernel_provider": "torch_compile",
"metric_name": "speed",
"metric_unit": "ms",
"gpu_name": "Ascend910B4",
"x_name": "V",
"x_label": "embedding dimension",
"x_values": [
1024,
2048,
4096,
8192,
16384,
32768,
65536,
131072
],
"y_values_50": [
0.1696000099182129,
0.20161999762058258,
0.20436999201774597,
0.20262999832630157,
0.20093999803066254,
0.21164000034332275,
0.23170000314712524,
0.23863999545574188
],
"y_values_20": [
0.16475999355316162,
0.1967879980802536,
0.19750800728797913,
0.19844000041484833,
0.19786399602890015,
0.21121999621391296,
0.2281000018119812,
0.23797999322414398
],
"y_values_80": [
0.17654000222682953,
0.20855599641799927,
0.2127159982919693,
0.20847199857234955,
0.20453999936580658,
0.21206000447273254,
0.2365799993276596,
0.23921999335289001
],
"timestamp": "2026-01-21 03:48:32",
"kernel_operation_mode": "forward",
"extra_benchmark_config_str": "{\"B\": 32, \"T\": 512, \"D\": 768, \"dtype\": \"torch.float32\"}",
"liger_version": "0.0.0"
},
{
"kernel_name": "embedding",
"kernel_provider": "liger",
"metric_name": "speed",
"metric_unit": "ms",
"gpu_name": "Ascend910B4",
"x_name": "V",
"x_label": "embedding dimension",
"x_values": [
1024,
2048,
4096,
8192,
16384,
32768,
65536,
131072
],
"y_values_50": [
61.94499969482422,
62.33565902709961,
61.86201858520508,
62.008819580078125,
62.232479095458984,
62.473838806152344,
63.21195983886719,
64.28494262695312
],
"y_values_20": [
61.94499969482422,
62.33565902709961,
61.86201858520508,
62.008819580078125,
62.232479095458984,
62.473838806152344,
63.21195983886719,
64.28494262695312
],
"y_values_80": [
61.94499969482422,
62.33565902709961,
61.86201858520508,
62.008819580078125,
62.232479095458984,
62.473838806152344,
63.21195983886719,
64.28494262695312
],
"timestamp": "2026-01-21 03:48:50",
"kernel_operation_mode": "backward",
"extra_benchmark_config_str": "{\"B\": 32, \"T\": 512, \"D\": 768, \"dtype\": \"torch.float32\"}",
"liger_version": "0.0.0"
},
{
"kernel_name": "embedding",
"kernel_provider": "huggingface",
"metric_name": "speed",
"metric_unit": "ms",
"gpu_name": "Ascend910B4",
"x_name": "V",
"x_label": "embedding dimension",
"x_values": [
1024,
2048,
4096,
8192,
16384,
32768,
65536,
131072
],
"y_values_50": [
1.5649499893188477,
1.5831799507141113,
1.6187798976898193,
1.6978700160980225,
1.8488800525665283,
2.2740001678466797,
2.995260000228882,
4.282050132751465
],
"y_values_20": [
1.5638200044631958,
1.581808090209961,
1.6171720027923584,
1.6966240406036377,
1.8473479747772217,
2.271728038787842,
2.993760108947754,
4.2774481773376465
],
"y_values_80": [
1.5658999681472778,
1.585103988647461,
1.6203559637069702,
1.6992720365524292,
1.8494240045547485,
2.2756519317626953,
2.9975199699401855,
4.283827781677246
],
"timestamp": "2026-01-21 03:49:03",
"kernel_operation_mode": "backward",
"extra_benchmark_config_str": "{\"B\": 32, \"T\": 512, \"D\": 768, \"dtype\": \"torch.float32\"}",
"liger_version": "0.0.0"
},
{
"kernel_name": "embedding",
"kernel_provider": "torch_compile",
"metric_name": "speed",
"metric_unit": "ms",
"gpu_name": "Ascend910B4",
"x_name": "V",
"x_label": "embedding dimension",
"x_values": [
1024,
2048,
4096,
8192,
16384,
32768,
65536,
131072
],
"y_values_50": [
1.5654499530792236,
1.579859972000122,
1.6195299625396729,
1.7007999420166016,
1.8484400510787964,
2.296339988708496,
3.0023000240325928,
4.276410102844238
],
"y_values_20": [
1.5641800165176392,
1.5789759159088135,
1.6186840534210205,
1.6993600130081177,
1.8475240468978882,
2.295095920562744,
3.000540018081665,
4.273176193237305
],
"y_values_80": [
1.5664000511169434,
1.5810760259628296,
1.6209640502929688,
1.7026759386062622,
1.8497200012207031,
2.298271894454956,
3.0036399364471436,
4.278992176055908
],
```

No source file or test is named. Start by locating the NPU embedding operator and its benchmark entry, then reproduce the Ascend910B4 forward and backward measurements; done means preserving accuracy while improving the reported performance gap against Hugging Face and torch_compile.

Written by the indexing model from the issue text.