PR #11444: [XLA:GPU] disable mask in cuDNN attention

Imported from GitHub PR openxla/xla#11444

1. cuDNN attention mask is not doing masking with -inf but multiply which is not correct. Hence disable patterns with mask.
2. Follow up PR to clean up the remaining mask related logic.
Copybara import of the project:

--
acf95b6cc7e1084026eaf87c0119ba3801ba8f8c by cjkkkk <ske@nvidia.com>:

disable mask

Merging this change closes #11444

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#11444 from Cjkkkk:remove_mask acf95b6cc7e1084026eaf87c0119ba3801ba8f8c
PiperOrigin-RevId: 624057479

third_party/xla/xla/service/gpu/cudnn_fused_mha_rewriter.cc

@@ -454,12 +454,6 @@ absl::StatusOr<bool> IsFlashAttention(
   bool is_flash_attention = is_seqlen_supported && is_hidden_dim_supported;
   if (!is_flash_attention) return false;
 
-  // TODO(hebecker): The fMHA rewriter is triggering some miscompile when used
-  // with cuDNN 8.9.6+. So this is temporarily disabling all the capabilities
-  // added by 8.9.6 and beyond:
-  cudnn_version =
-      std::min(cudnn_version, stream_executor::dnn::VersionInfo(8, 9, 5));
-
   // going backwards to check compatibility
   if ((is_training && (s_q < 64 || s_kv < 64)) &&
       !IsComputeCapabilityAndCudnnSupported(
@@ -1865,6 +1859,10 @@ absl::StatusOr<bool> CudnnFusedMHARewriter::Run(
       if (!matched_result.has_match) {
         continue;
       }
+      // disable cuDNN mask input
+      if (matched_result.matched_mask) {
+        continue;
+      }
       // We check the validity of bmms here before canonicalization so we don't
       // modify the graph if mha fusion is not possible
       // Relax 512 constraint if it is flash attention