[ROCm] Fix LoadKernel to use refcounted module path for proper cleanup

[magaonka-amd]

Motivation

ROCm's RocmExecutor has two kernel/module loading paths with
different caching behavior:

1. LoadModule → LoadModuleFromHsaco → populates both
   gpu_binary_to_module_ (refcounted) and in_memory_modules_
2. LoadKernel → directly inserts into in_memory_modules_ only,
   bypassing gpu_binary_to_module_ entirely

When UnloadKernel calls UnloadGpuBinary, it only looks in
gpu_binary_to_module_ for cleanup. Since LoadKernel never populated
that map, the cleanup was a no-op: in_memory_modules_ entries and
loaded hipModule_t objects were never removed, leaking until
executor destruction.

This caused stale module cache entries when CustomKernelThunk's
owned HSACO buffers were freed and reallocated at the same address.
The cache returned old modules that didn't contain the expected
kernels, producing flaky hipErrorNotFound failures in sort and
LuSolve tests under xdist parallelism.

Fix by routing LoadKernel through LoadModuleFromHsaco, so every
loaded module participates in the refcount mechanism and is properly
cleaned up when the last kernel referencing it is destroyed.

With this fix properly wiring LoadKernel through the
refcounted LoadModuleFromHsaco path, stale cache entries are cleaned
up before address reuse can occur. The content-hash key introduced
in PR openxla#40419 is no longer needed as a workaround.

Revert to using raw pointer identity as the ModuleHandle key
while keeping the hipPeekAtLastError
guard in GetModuleFunction for diagnostic purposes.

[claude]

Claude Code Review Summary

Clean PR — no issues found.

This change correctly fixes a resource leak in RocmExecutor::LoadKernel by routing through the refcounted LoadModuleFromHsaco path instead of directly inserting into in_memory_modules_. The pattern now matches the CUDA executor's LoadKernel/LoadModuleFromCuBin approach exactly. The revert of the content-hash ModuleHandle key is sound since the root cause (missing refcount path) is now addressed, making raw pointer identity sufficient again.

Lock acquisition order, map access patterns, and BUILD dependency cleanup are all correct.

🤖 Reviewed with Claude Code

[magaonka-amd]

I'm inclined to keep this change from my original PR hence did partial revert here. I think it is useful to stop execution if kernel is not found else get ourselves into undefined behavior which is much harder to debug.
https://github.com/openxla/xla/pull/40419/changes#diff-41b3141e3358c4655366766128707b51ed7ad7ec0d7f7f0059770c01fced1859R179

but I'm open to suggestions on this.

[draganmladjenovic]

I'm inclined to keep this change from my original PR hence did partial revert here. I think it is useful to stop execution if kernel is not found else get ourselves into undefined behavior which is much harder to debug. https://github.com/openxla/xla/pull/40419/changes#diff-41b3141e3358c4655366766128707b51ed7ad7ec0d7f7f0059770c01fced1859R179

but I'm open to suggestions on this.
Remove it. It makes no sense to spray code with hipPeakAtLastError.

[magaonka-amd]

Okay Dragan removed and pushed PR upstream thanks for feedback.
openxla#40847 pushed fix upstream with new fix and revert of my old fix.

[draganmladjenovic]

LGTM. It seem to match what cuda does.