[XLA:GPU] Allow cuDNN scaled dot fusions in the gemm autotuner

[sergey-kozub]

📝 Summary of Changes Allow selecting cuDNN gemm configs when autotuning scaled dot fusions.

🎯 Justification cuDNN has a kernel for block scaled dot operations, this PR enables it in the autotuner. Note: XLA flag --xla_gpu_experimental_scaled_dot_with_triton is required to enable this.

🚀 Kind of Contribution ✨ New Feature ⚡️ Performance Improvement

[golechwierowicz]

Following tests fail

• CollectiveOpsTestE2EShardedUnsharded.BlockScaledDotNonContractingAndContracting on B200.
• GemmFusionAutotunerLevelSweep/GemmFusionAutotunerLevelTest.Deviceless/3 on H100.

Can you please fix?

[sergey-kozub]

GemmFusionAutotunerLevelSweep/GemmFusionAutotunerLevelTest.Deviceless/3

Fixed the issue breaking this test.

CollectiveOpsTestE2EShardedUnsharded.BlockScaledDotNonContractingAndContracting

This test fails for me on HEAD (i.e. seems unrelated to this PR), could you please confirm? (whether it passes for you on HEAD).

[golechwierowicz]

This test fails for me on HEAD (i.e. seems unrelated to this PR), could you please confirm? (whether it passes for you on HEAD).

It passes for us at HEAD.

[sergey-kozub]

It passes for us at HEAD.

Looking into this.

[golechwierowicz]

And also please make sure you include the right build deps

to fix run:
	build_cleaner... /xla/service/gpu/transforms:block_scaling_rewriter_test
   24 | #include "third_party/tensorflow/compiler/xla/hlo/ir/hlo_casting_utils.h"
      |          ^
to fix run:
	build_cleaner... /xla/service/gpu/transforms:block_scaling_rewriter_test
   26 | #include "third_party/tensorflow/compiler/xla/hlo/testlib/filecheck.h"