[Bug] test_proxy_split_k_heuristic fails on RTX 5090 (170 SM)

## Summary `tests.msa_ops.test_proxy_fp4.test_proxy_split_k_heuristic` fails on an RTX 5090 (170 SMs): ``` actual split_k = 40 expected split_k = 43 ``` The test expectation is derived from the current closed-form SM scaling, while `_split_k_makespan_argmin` selects 40 from the modeled makespan. ## Evidence For the failing input, the modeled makespan is: - split_k 40: 14 - split_k 41: 14 - split_k 42: 14 - split_k 43: 26 So 40 is a valid first argmin and 43 is not an argmin for the current cost model. This looks like a test expectation/heuristic-contract mismatch rather than a driver or hardware fault. ## Environment - GPU: RTX 5090 - SM count: 170 - Test: `tests.msa_ops.test_proxy_fp4.test_proxy_split_k_heuristic` - CUDA 13.0 local run: [FI-Test-Runner-Docker #22](nv/view/TRT-Perf-Test-pipeline/job/FI-Test-Runner-Docker/22/) - CUDA 12.9 local run: [FI-Test-Runner-Docker #23](nv/view/TRT-Perf-Test-pipeline/job/FI-Test-Runner-Docker/23/) ## Suggested resolution Please clarify the intended contract: 1. If the heuristic should return the first makespan argmin, update the test expectation for 170 SMs to 40. 2. If 43 is required, update the makespan/candidate-selection logic and document the tie-breaking rule. This issue is intentionally filed in FlashInfer rather than as an NVBug because the observed value is supported by the current FlashInfer makespan model.