[DevTSAN] Support detecting data race for local memory

[zhaomaosu]

• Add builtins _tsan_cleanup[static|dynamic]_local to initialize local shadow
• Add missing lock for kernel launch api

[Copilot]

Pull Request Overview

This PR extends ThreadSanitizer support to detect races in local (shared) memory by allocating and initializing a separate shadow region per workgroup and passing local-argument metadata through the interceptor and runtime.

• Added AllocLocalShadow API to ShadowMemory and provided CPU/GPU implementations.
• Enhanced the interceptor and DDI to record local kernel arguments and calculate per-kernel local shadow offsets.
• Updated the device runtime (tsan_rtl.cpp), SPIR-V instrumentation, and end-to-end tests to initialize, clean up, and report races on local memory.

Reviewed Changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 2 comments.

File	Description
tsan_shadow.hpp & tsan_shadow.cpp	Defined AllocLocalShadow() and implemented it for CPU/GPU shadows.
tsan_interceptor.hpp & tsan_interceptor.cpp	Recorded local arg sizes, computed workgroup counts, and invoked AllocLocalShadow.
tsan_rtl.cpp	Mapped local memory to shadow space, added cleanup hooks, and marked local races.
llvm/.../ThreadSanitizer.cpp	Instrumented static/dynamic local memory in SPIR-V and factored out common utils.

Comments suppressed due to low confidence (2)

unified-runtime/source/loader/layers/sanitizer/tsan/tsan_interceptor.hpp:148

• [nitpick] Constructor parameter LocalWorkSize shadows the member of the same name; consider renaming the parameter (e.g. pLocalWorkSize) to avoid confusion.

             const size_t *GlobalWorkSize, const size_t *LocalWorkSize,

unified-runtime/source/loader/layers/sanitizer/tsan/tsan_report.cpp:30

• [nitpick] The updated warning message only indicates memory type and omits the actual race address, which may reduce debugging clarity; consider including the address in the log string.

           "====WARNING: DeviceSanitizer: data race on {}",

[yingcong-wu]

LGTM with one nit question.

[zhaomaosu]

Hi @intel/unified-runtime-reviewers, could you please help review this PR? Thanks.

[zhaomaosu]

Hi @intel/llvm-gatekeepers, this PR is ready to be merged. Thanks.