Waitcheck GEMM: resolve each hipBLASLt solution to its real object/index/entry (no fabricated copies)

Follow-up from the PR #324 review (BLOCKER: "Whole-object scans are presented as selected GEMM checks").

scripts/sanitizers/prepare_gemm_isa.py currently copies one unbundled code object under every sol_<idx>.hsaco filename, so distinct hipBLASLt solutions are fabricated as identical objects. Waitcheck then scans the whole object.

Already landed in PR #324 (safe, in-repo parts):

• Waitcheck no longer attributes object-scan findings to a synthetic GEMM kernel name (kernel_name=None for scan identities).
• Identical scan identities now produce a single object-level invocation (dedup by code object).

Remaining (needs real hardware + hipBLASLt metadata):

• Resolve each selected solution to its real code object, image index, and kernel entry offset so kernel-scoped Waitcheck can run exact-entry per solution.
• Introduce an explicit object-scope result shape (no kernel identity/name) for the genuine whole-object-scan fallback, instead of reusing a kernel-named identity.
• Update prepare_gemm_isa.py to stop emitting fabricated per-solution copies.

Test: identical scan identities must produce one object-level invocation and no fabricated kernel attribution (unit test landed); add hardware coverage for real per-solution exact-entry resolution.

Start with scripts/sanitizers/prepare_gemm_isa.py and the Waitcheck changes from PR #324. Trace how selected hipBLASLt solutions currently map to sol_.hsaco, then use the available metadata and real hardware to resolve each solution's code object, image index, and kernel entry offset. Done means exact-entry checks use real per-solution objects, genuine whole-object scans have an explicit object-scope result, and fabricated copies are no longer emitted.