Add cuDNN tensor IR support

[cerisier]

Summary

• Expose cudnn_engines_tensor_ir and cudnn_engines_tensor_ir_static through cuDNN 9 BUILD/proxy targets.
• Gate tensor IR on cuDNN >= 9.21.0 and actual archive contents, so cuDNN 9.20 redists keep building without the new library.
• Add e2e coverage at the introduction boundary by pinning the CUDA 13.3 smoke case to cuDNN 9.21.0 and checking the shared/static tensor IR labels.

Context

This mirrors the tensor_ir support being added in rules_ml_toolchain, but keeps the target conditional because libcudnn_engines_tensor_ir first appears in cuDNN 9.21.

Checked locally with bazel build //..., plus CUDA 13.3/cuDNN 9.21 tensor IR and CUDA 13.2/cuDNN 9.20 cuDNN smoke builds.