Update trtllm-gen FMHA cubins to fix context SWA page-skip

Update TRTLLM_GEN_FMHA artifact path and checksum to pick up cubins
built from trtllm-gen with the context SWA page-skip fix. The new cubins
skip loading out-of-window KV pages in context (prefill) kernels,
preventing NaN corruption from null blocks in the KV cache.

Fixes: https://nvbugspro.nvidia.com/bug/5922676

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: PerkzZheng

flashinfer/artifacts.py

@@ -135,7 +135,7 @@ class ArtifactPath:
     When compiling new cubins for backend directories, update the corresponding path.
     """
 
-    TRTLLM_GEN_FMHA: str = "55bba55929d4093682e32d817bd11ffb0441c749/fmha/trtllm-gen/"
+    TRTLLM_GEN_FMHA: str = "84ee55b93ba92b1f64bbf727fd1cf38d9d3058a7/fmha/trtllm-gen/"
     TRTLLM_GEN_BMM: str = (
         "39a9d28268f43475a757d5700af135e1e58c9849/batched_gemm-5ee61af-2b9855b/"
     )
@@ -155,7 +155,7 @@ class CheckSumHash:
     """
 
     TRTLLM_GEN_FMHA: str = (
-        "f2c0aad1e74391c4267a2f9a20ec819358b59e04588385cffb452ed341500b99"
+        "99add183a75a55a0aa77c6f61fa58cd6a2b40709effbdf8527e0a21588cdc7c9"
     )
     TRTLLM_GEN_BMM: str = (
         "db06db7f36a2a9395a2041ff6ac016fe664874074413a2ed90797f91ef17e0f6"