[BUG] Blackwell MLA perf for split-k

Describe the bug

I want to pass variable split_k via Arguments.ptr_split_kv on-device. I need to also set max value of split k Arguments.split_kv. The issue is that I need to keep max value of split k static (like 16 or so), but as soon as I set it higher than 1 perf degrades a lot (I pass only 1 to ptr_split_kv). This makes ptr_split_kv on-device not practically usable.

Another data point. Perf degrades 2x if I just split by 4x. This is really not expected as it's not doing 2x work (yes, there is some overhead, but it should be much smaller).

$ /77_blackwell_mla_2sm_fp16 --b=512 --k=2048 --page=128 --verbose --split_kv=1
###### B 512 MLA H 128 D_rope 64 D_latent 512 Q 1 K 2048 Gen None Split 1 Gen None #SM 148
 [--] 128x128 fp16 persistent          : 1073.23 TFLOPS/s 4.13682 TB/s
       t=326.475 us, smem=210944b
 [--] 128x128 fp16 individual          : 1071.13 TFLOPS/s 4.12872 TB/s
       t=327.115 us, smem=210944b

$ /77_blackwell_mla_2sm_fp16 --b=512 --k=2048 --page=128 --verbose --split_kv=4
###### B 512 MLA H 128 D_rope 64 D_latent 512 Q 1 K 2048 Gen None Split 4 Gen None #SM 148
 [--] 128x128 fp16 persistent          : 547.292 TFLOPS/s 2.10956 TB/s
       t=640.213 us, smem=210944b
 [--] 128x128 fp16 individual          : 508.058 TFLOPS/s 1.95833 TB/s
       t=689.653 us, smem=210944b

Steps/Code to reproduce bug
See description for repro.

Expected behavior
I expect perf not to degrade significantly when max split k is increased.

Environment details (please complete the following information):
B200

Additional context
Add any other context about the problem here.

Start with the 77_blackwell_mla_2sm_fp16 entry point and reproduce the two commands using --split_kv=1 and --split_kv=4 on B200. Trace how Arguments.ptr_split_kv and Arguments.split_kv are handled, then measure whether raising the maximum split-k value leaves the split_kv=1 case near its baseline without the reported performance regression.