#1388 hip: gfx1151 hipBLASLt tuning table for setup's ROCm 7.14.0a20260608 (hipBLASLt 1.4.0)

Description

Follow-up to #1267. Since 0.1.40.2, setup installs the ROCm 7.14.0a20260608 wheels for gfx1151. Their hipBLASLt is
**1.4.0** (`100400`), which neither shipped gfx1151 table matches (`100401`, `100500`), so a fresh setup reads the
prompt through plain hipBLAS:

```
[!]  no hipBLASLt tuning table for gfx1151 with hipBLASLt 100400 (have: gfx1151-hipblaslt-100401.txt, gfx1151-hipblaslt-100500.txt): the prompt's dense matrix products use plain hipBLAS
```

This adds `tools/hip/gfx1151-hipblaslt-100400.txt`, calibrated against that library, and a docs entry.

**Calibration.** Ryzen AI Max+ 395 (Radeon 8060S, gfx1151, 128 GB), Fedora 44, kernel 7.2.8, a fresh
`./setup.sh --backend hip` of 0.1.40.2 (e8ca9af), `rocm-sdk-core 7.14.0a20260608`. `tune_hipblaslt` built from that
tree against setup's wheels, run on the 90 shapes of `gfx1151-hipblaslt-100401.txt` (the prompt shapes at
`--prefill 16384` included):

```sh
CASES=$(awk 'NR>2 {printf " --case %s,%s,%s,%s,%s", $1, $5, $2, $3, $4}' tools/hip/gfx1151-hipblaslt-100401.txt)
./build-hip/tune_hipblaslt --workspace-mib 32 $CASES --tuning-out gfx1151-hipblaslt-100400.txt
```

90 rows, the same geometries as the 100401 table. Per shape 3.1x to 14.1x faster than plain hipBLAS (median 6.2x);
every chosen solution passed the tool's accuracy check.

**Tests**, with `STRATA_HIPBLASLT_TUNING` pointing at the new table:

- `hip_prefill_hipblaslt_gemm`: PASS, `hipBLASLt tuning enabled (90 rows, gfx1151, version 100400)`,
  `launches=4 fallbacks=0`
- `hip_prefill_hcd_exact_parity`: PASS, solutions 1176 / 1177 at 9 chunk sizes (T 64 to 16384), 0 outputs differ,
  `fallbacks=0`
- HIP ctest: 61 of 62 pass; `ple_parity` fails only because the Q2_0 model fixture is not on this machine

setup picks the table up with no code change (`hipblaslt_version()` reports 100400 for these wheels, and
`hipblaslt_table("gfx1151", ...)` returns the new file).

Not measured yet: the end-to-end prompt speed of a whole model with this table.

🤖 Generated with [Claude Code](https://claude.com/claude-code)