fix: use the prompt cache for Gemma 4 text requests on VLM loads

[solderzzc]

Fixes #200

Problem

Gemma 4 26B A4B auto-loads as a VLM, and every VLM load skipped the prompt cache, so each text turn re-prefilled its whole conversation.

Changes

• Use the prompt cache for text requests on MLXVLM.Gemma4 (its text-only path is stateless). Qwen-VL and the other VLM/Omni models stay skipped, since they need the LMOutput state (ropeDeltas) of the cached prefix. Video inputs now also count as multimodal.
• Slice the input along the sequence axis, keeping the [1, T] rank and the mask (an axis-0 slice cut the batch axis on a hit).
• Detach RotatingKVCache snapshots on save and restore: decode steps write into the ring buffers in place, which corrupted saved snapshots.
• For caches with sliding-window layers, snapshot synchronously at the turn boundary (last <|turn> / <|im_start|>) instead of one decode token late. A wrapped ring that would need a trim deeper than one slot is treated as a miss.
• onPrefillDone no longer saves for requests that skip the cache (it previously also saved multimodal prompts).
• Replace the tautological skip-predicate tests with the real shouldSkipPromptCache function; add ring-buffer, slicing and boundary tests.

Behaviour change

Non-Gemma LLMs with a wrapped sliding-window cache (--ctx-size) may now miss where the old late save produced a misaligned snapshot.

Testing

• PromptCacheTests + PromptCacheRotatingTests: 23 tests pass. With the detach disabled, two of the ring tests fail.
• gemma-4-e2b-it-4bit and gemma-4-26b-a4b-it-4bit, ~2.5k / ~4.3k token system prompt, two turns, temperature 0: HIT on repeat and on turn 2, and the turn-2 output equals a cold run on a fresh server.
• gemma-4-e2b with --audio: image request x2, audio request x2, then text x2. The only HIT is the repeated text request; nothing hits after an image or audio request.
• Qwen2.5-0.5B-Instruct-4bit with --ctx-size 2048 (RotatingKVCache), ~1.8k-token prompt (ring not wrapped): HIT on repeat and on turn 2, and the turn-2 output equals a cold run.
• Same model, ~2.5k-3k-token prompt (ring wrapped): HITs occur, but the cached turn-2 output differs from a cold run. A cold run alone also changes with --prefill-size (256 vs 512/1024), so this is the sliding window's sensitivity to prefill chunking rather than a restore error. This was inferred from outputs, not by comparing logits.
• CI: build_and_unit_test and the full integration matrix pass.

AI disclosure

Written with AI assistance (Claude Code, Claude Sonnet 5.5) and reviewed by AI agents; the author has not yet reviewed it line by line.

• I have read this PR description

🤖 Generated with Claude Code