#2303 hip: a hipBLASLt tuning table for gfx1103 (Radeon 780M)

Descripción

Issue: Follow-up to #1423

## Summary

The Radeon 780M (gfx1103, RDNA3, 12 CUs) builds and runs (issue #1423), but there was no hipBLASLt table for it, so
the prompt's dense GEMMs ran on plain hipBLAS. This adds `tools/hip/gfx1103-hipblaslt-100401.txt`, calibrated on one
780M with ROCm 10.1's hipBLASLt 1.4.1 (version number `100401`; the same version number as the shipped gfx1150
table, which was calibrated on 7.14.1).

## What changed

- `tools/hip/gfx1103-hipblaslt-100401.txt` (new, 143 rows): the 16 dense GEMM geometries a Qwen3.8-Flash-Next
  prompt logs at T = 64, 128, ..., 16384, calibrated with `tools/hip/tune_hipblaslt --workspace-mib 32`. Two passes
  back to back; each row keeps the solution with the lowest median of all six repetitions, and a row is left out
  when no candidate beats plain hipBLAS, so 143 of the 144 cases are in: 1.05-23.30x per GEMM, geometric mean 4.95x.
  The one left out (`bf16 N=1 K=2560 T=64`) falls back to hipBLASEx.
- `docs/AMD_HIP.md`: a "Shipped tables" entry.
- `bench/results/2026-10-12-community-gfx1103/`: the rig, the method, the runs (`runs.csv`), the server config and
  the raw logs.
- `bench/results/COMMUNITY.md` and `docs/COMMUNITY_BENCHMARKS.md`: one entry each.

## Extra Notes

- Prompt speed on the 780M (Q2_0): **90 tok/s on plain hipBLAS -> 187 with the table** at a 4K prompt, medians of an
  interleaved A/B; with the gfx1103 exact switches and the WMMA opt-ins, 277 at 4K and 272 at 32K (the earlier
  bench configuration; the shipping one adds `--spec 2 --lookup-chain 4`, which measures 276.7 at 4K).
- Context ladder to 131K on the shipping configuration (fresh prompts from this repository's own docs): prefill
  holds ~310 tok/s to 32K, then 258.2 at 64K, 215.5 at 96K and 198.5-221.4 at 128K across two runs (that is the
  full-read average; the last ~14K tokens at 128K depth read at ~206 tok/s with the prefix cached). Decode 16 ->
  ~10 tok/s with depth; GTT stays at 38.1 of 50 GiB.
- `hip_prefill_hipblaslt_gemm` passes with it (`launches=4 fallbacks=0`); `hip_prefill_hcd_exact_parity` skips,
  because the table picks no 1176 / 1177 for the HC down GEMM - the gfx1150 table gets the same skip, which is why
  `STRATA_HCD_EXACT` is not in the gfx1103 switch set.
- setup installs for gfx1103 only with `STRATA_EXPERIMENTAL_GFX1103=1`, and then picks the table up when the
  installed hipBLASLt reports 1.4.1.
- Not tested: another hipBLASLt build (the engine refuses other version numbers), other models, contexts above 131K.