[samuelishida]

Adds tools/hip/gfx1100-hipblaslt-100202.txt, calibrated with the branch's own tuner on an RX 7900 XTX under ROCm 7.2.1.

• 26 of the 30 covered shapes validated.
• Four shapes at T=4096 produce non-finite output in every hipBLASLt solution on 100202; those rows are intentionally absent and the runtime falls back to hipblasGemmEx.

Also makes hip_prefill_hipblaslt_gemm self-locate a matching shipped table for the runtime's hipBLASLt version (falling back from STRATA_HIPBLASLT_TUNING), so the parity test runs unconditionally on a machine with a committed table. Documents the table and its fallback behavior in docs/AMD_HIP_PERFORMANCE.md.

[Niko1221]

Thanks @samuelishida. Two things for a new PR: the test uses STRATA_TUNING_TABLE_DIR, which isn't defined in CMake (the HIP test build breaks), and the doc implies the engine finds the table itself; only the test does. Also note our setup currently takes the highest-numbered gfx1100 table regardless of the installed hipBLASLt version; we'll fix setup to pick the table that matches the runtime, and then a 100202 table can go in cleanly. We're closing PRs that wait on changes so the list stays manageable; this is not a no. Please open a new PR when it's ready and link this one.