chore: extract MTP benchmark tooling from #109

[solderzzc]

#109 bundled three unrelated things: a sliding-window KV cap, a Gemma4-specific MTP-assistant auto-resolution wiring, and this benchmark tooling. Only the tooling comes forward here.

Why not the rest of #109

• The KV cap was independently ported and evaluated earlier this session — no measurable benefit at 150 tokens or ~9k context. Parked, not because of a rebase conflict, on the numbers.
• The auto-resolution wiring (`Gemma4MTPRegistry`, an `mtpAsstRef` reusing the draft-model speculative path) duplicates what feat: make --mtp work for model families that ship MTP heads separately #137 already shipped, more generally, as the explicit `--mtp-assistant-model` flag. A literal rebase would land a second, Gemma4-only implementation of a feature that already works for any `DualModelMTP` architecture — worse, not newer.

Why this part rebases clean

`mtp_bench.py` drives the server with `--mtp --num-mtp-tokens N --turbo-kv` — the single-checkpoint MTP path, not the Gemma4 dual-model wiring that's in conflict. Verified none of these three files reference `Gemma4MTPRegistry`, `--mtp-assistant-model`, or anything else from the withheld part of #109. All three flags it uses already exist on main independently of that wiring.

`run_benchmark.sh`'s Test 13 previously shelled out to `swift run Gemma4MTPBench` — that product no longer exists in `Package.swift`, so the path was already broken on main before this PR. It now drives `mtp_bench.py` against the regular `SwiftLM` binary.

On the README numbers

Carried over from the original PR's measurements, not reproduced in this extraction — a 40K/100K context run is multi-hour. The 8-bit table matches data already used as reference in mlx-swift-lm#46 and the #137 comment thread earlier this session, so it isn't new to this repo's history, just newly landing in the README.

Verification

• `run_benchmark.sh` — syntax checked
• `mtp_bench.py` — compiles, `--help` parses correctly
• confirmed `--mtp`, `--num-mtp-tokens`, `--turbo-kv` all exist on main independent of `--mtp-assistant-model`

Refs #109

🤖 Generated with Claude Code