[NVIDIA] Relax previous restraints for using flashinfer gdn decode

[kaixih]

The combination of --linear-attn-decode-backend flashinfer and --mamba-scheduler-strategy no_buffer was previously blocked with a hard ValueError in server_args.py. The root cause was a bug in the FlashInfer bf16 CuTe-DSL kernel (SM100+/Blackwell): CUDA graph decode uses padded batches where padding slots have initial_state_indices = -1, but the kernel had no guard for negative indices. Now the flashinfer side has fixed it (flashinfer-ai/flashinfer#2810, and later flashinfer-ai/flashinfer#2679).

Also added tests and got the results below on 8xB200:

============================================================
Qwen3.5-397B-A17B-NVFP4 Results Summary
Dataset: gsm8k
Baseline: 0.95
============================================================

Model 1: nvidia/Qwen3.5-397B-A17B-NVFP4
  Accuracy: PASS
  Score: 0.990

Model 2: nvidia/Qwen3.5-397B-A17B-NVFP4
  Accuracy: PASS
  Score: 0.990

Model 3: nvidia/Qwen3.5-397B-A17B-NVFP4
  Accuracy: PASS
  Score: 0.995

============================================================
OVERALL: ALL TESTS PASSED
============================================================

.
----------------------------------------------------------------------
Ran 1 test in 940.944s

OK

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[kaixih]

cc. @Fridge003 @hlu1

[kaixih]

Sorry, this has already fixed the issue: #21861.

Closing.