Potential numerical accuracy issue in the fused_moe implementation

Test environment: CTK 13.1, torch = 2.9.1+cu130, cuda-tile = 1.0.0, single B200 GPU

When using a parameter set (num_tokens, hidden_size, moe_intermediate_size, n_experts, top_k) = (128, 4096, 2048, 16, 4) in tests/ops/test_moe.py and add torch.manual_seed(0) in line 101, will have mismatched results:

(Earlier output omitted)
>       assert passed, f"\n{failed_msgs}"
               ^^^^^^
E       AssertionError:
E               *** OUTPUT 0 DID NOT MATCH THE REFERENCE (rtol=0.1, atol=0.1) ***
E                       allclose: False
E                       matched: 523466 / 524288 [99.84%]
E                       ref range:    -1.1700e+02 :  1.2300e+02
E                       test range:   -1.1650e+02 :  1.2300e+02
E                       |ref| range:   0.0000e+00 :  1.2300e+02
E                       |test| range:  0.0000e+00 :  1.2300e+02
E                       max absolute difference:  1.0000e+00
E                       max relative change:      4.3000e+01
E                       max max mean change:      2.0000e+00
E                       max arith mean change:    5.1400e+02
E                       shape: torch.Size([128, 4096]) stride: (4096, 1) dtype: torch.bfloat16
E                       mismatched indices:tensor([[   0, 1842],
E               [   0, 2071],
E               [   1,  675],
E               ...,
E               [ 127, 1728],
E               [ 127, 2207],
E               [ 127, 2628]])

For the same parameter set, if dtype is changed to float16, the test can pass.
Using a smaller hidden_size or a smaller moe_intermediate_size can also pass the test.

Start with tests/ops/test_moe.py, reproduce the reported parameter set with torch.manual_seed(0), and compare bfloat16 against float16. Trace the fused_moe entry point used by that test and investigate the mismatched outputs for the larger dimensions. Done means the bfloat16 case matches the reference within the test tolerance without regressing the passing cases.