#755 hip: gfx1100 hipBLASLt 1.5.0 tuning table (ROCm 10.2 nightly) — closes the gfx1100 100500 gap

Beschreibung

## Summary

Ship `tools/hip/gfx1100-hipblaslt-100500.txt` — a hipBLASLt tuning table for the RX 7900 XTX calibrated against hipBLASLt **1.5.0** (version number 100500), the same gap #481 closed for gfx1201. This is an engine-tree companion to the results-only #745 (its second commit measures exactly this table; per-run data lives there).

**Why:** the engine refuses a table whose version doesn't match the installed hipBLASLt, so on a ROCm 10.2 nightly (hipBLASLt 1.5.0) the shipped `gfx1100-hipblaslt-100100/100200.txt` are refused and prefill silently runs on plain hipBLAS. Every gfx1100 nightly user is on the untuned path today.

**What:** the same 16 dense GEMM geometries as the other gfx1100 tables (bf16 + f16, T = 4096 and 8192 → 32 rows), calibrated with `tune_hipblaslt` run against that exact nightly library (ROCm 10.2.0a20261003 SDK, `libamdhip64.so.7.17.26392`; every candidate gated at rel-L2 ≤ 1e-4 / max-abs ≤ 1e-2 / no padding writes; each row is its case's measured best). `setup` picks it up by name once the installed hipBLASLt reports 100500 — no config change needed.

**Measured** (0.1.38-lineage nightly binary of this tree, Qwen3.8-Flash-Next IQ3_S greedy, arms differing only in `STRATA_HIPBLASLT_TUNING`, median of 3 clean runs each):

| prompt tokens | with table | plain hipBLAS | gain |
| ---: | ---: | ---: | ---: |
| 131071 | 1687 tok/s | 926 tok/s | **1.82x** |

Decode unchanged (62.3 vs 60.7 tok/s, overlapping ranges — the table is prefill-only, as designed). Per-run data, calibration command and tuner log: [bench/results/2026-10-04-community-rx7900xtx-hipblaslt-100500/](https://github.com/Niko1221/Strata/pull/745) in #745.

**docs/AMD_HIP.md:** one entry in the shipped-tables list, mirroring the gfx1201 1.5.0 entry (including the caveat that the version number is the only thing the engine can check — recalibrate with `tune_hipblaslt` for a different 1.5.0 build).

## Checklist

- Results-only benchmark evidence is in #745; this PR is the engine-tree piece, kept separate per `docs/COMMUNITY_BENCHMARKS.md`.
- Hardware: RX 7900 XTX 24 GB (gfx1100), Ryzen 9 7900X, Fedora 44; same host as #745.