tools : split Metal FA-vec tuning into a standalone tuner

[forforever73]

Overview

cont #25750, implementing two of its follow-ups.

The (Q, NE) sweep / compression / emit generator moves out of tests/test-backend-ops.cpp into a new binary ggml-metal-tuning (tools/tuning), which links only ggml and is gated on GGML_METAL, and gains a closed-loop cooldown for thermal drift. What stays in the test file is a 240-case Metal-gated numerical slice that every Apple-silicon CI run executes.

ggml-metal-tuning fa-vec [-b MTL0] [--dtype f16,q4_0] [--dk 128,192]
                         [--reps 7] [--seed 1234]
                         [--no-cooldown] [--cool-drift 0.10] [--cool-eps 0.03]
                         [--cool-max-wait 120] [--cool-max-retry 2]

stdout is the pasteable table, stderr is everything else (> rows.txt 2> sweep.log). The old generator put both on one stream, so the rows had to be dug out of six blocks in a 2396-line file, and the 2>/dev/null it documented discarded the Metal log; keeping that log instead tore 4 of 108 cell lines in half per run.

Where the old sweep printed a drift warning and kept the measurement, the cooldown discards every candidate timed since the last clean anchor, waits for the GPU to cool, re-measures them, and drops the cell from the table if it never converges, so a throttled timing can no longer reach the table.

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: Yes, claude was used for the port, the cooldown implementation and the validation runs. The decisions (what stays in the test file, the dirty-window definition, dropping untrusted cells rather than pausing) were made by a human.