#170 hip: gfx1100 hipBLASLt tuning table for ROCm 7.2.1 (100202)

描述

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`.