fix: MLLM hybrid batching + message normalization

[Thump604]

Summary

Split from #165 — prefix cache hybrid changes deferred to #217 (better approach).

This PR contains the non-prefix-cache fixes from #165:

• mllm_batch_generator.py: Hybrid cache handling (ArraysCache + KVCache interleaved) for _make_batch_cache, merge, filter, extract, extend operations
• mllm_scheduler.py: Hybrid cache scheduling for BatchedEngine
• server.py: _normalize_messages() — maps developer -> system role, merges consecutive same-role messages. Applied to both MLLM and LLM paths.
• tokenizer.py: VLM tokenizer loader with mlx-lm fallback
• qwen3_5_mllm.py: Qwen3.5 MLLM patch for hybrid batching

Fixes #137 (developer role crash on Qwen 3.5 templates). Fixes OpenCode crash when sending [system, system, user, user] message sequences.

Test plan

• 524-line test suite for hybrid cache operations (test_mllm_hybrid_cache.py)
• Validated on M2 Ultra 128GB with Qwen3.5-122B-A10B (hybrid MoE, ArraysCache + KVCache layers)
• Tested with BatchedEngine continuous batching + prefix cache enabled

Context

The original #165 also included prefix cache hybrid support (prefix_cache.py). That part overlaps with #217 which has a cleaner storage-type dispatch model. This PR extracts the unique, non-overlapping fixes so they can be reviewed and merged independently.

[Thump604]

Evidence from M2 Ultra 128GB, Qwen3.5-122B-A10B-VLM-MTP-5bit, BatchedEngine:

Test	Result
Text request through BatchedEngine	PASS
Message normalization (consecutive user messages)	PASS
Developer role mapping	PASS

All 3/3 tests pass. This is the slim split of #165 with prefix cache changes deferred to #217.

The message normalization fix is particularly important -- Qwen 3.5 templates reject consecutive same-role messages and system messages out of order. Without it, multi-turn agent conversations crash (OpenCode sends [system, system, user, user] sequences, and the OpenAI Responses API sends developer role instead of system). This has caused production crashes three times.

[Thump604]

@janhilgard - this fixes multi-turn agent crashes from developer role messages and consecutive same-role messages. Would appreciate a review if you have time.

[Thump604]

Split into three focused PRs for easier review:

• fix: normalize messages before chat template application #240 - Message normalization (_normalize_messages, developer role mapping, consecutive same-role merge)
• fix: MLLM hybrid model batching (ArraysCache support) #241 - MLLM hybrid batching (ArraysCache support, _make_batch_cache rewrite, extend fix, error handling)
• fix: MLLM scheduler streaming detokenizer + VLM model pre-detection #242 - MLLM scheduler streaming detokenizer + VLM pre-detection

Each PR is based on origin/main and stands alone with no cross-dependencies.