[Fix] Re-encode multimodal embeddings after cache mismatch

[mickqian]

Motivation

based on #32404

A compact multimodal feature-hash collision can return a cached embedding whose token count does not match the placeholder span in the request. The mismatch currently reaches _adjust_embedding_length and aborts prefill with Insufficient multimodal embedding length.

Modifications

• Validate cached embedding token counts against the item offsets on all chunked-prefill cache-hit paths.
• Evict malformed entries and re-encode them instead of passing them to the scheduler.
• Include (hash, token_count) in the per-batch temporary map so colliding items with different placeholder lengths are not deduplicated.
• Add regression coverage for full, per-item, and batched cache paths.

Validation

• pre-commit run --files python/sglang/srt/managers/mm_schedule.py test/registered/chunked_prefill/test_mm_chunked_embedding_unit.py
• python3 -m compileall -q python/sglang/srt/managers/mm_schedule.py test/registered/chunked_prefill/test_mm_chunked_embedding_unit.py
• Pytest was unavailable locally because PyTorch/pytest are not installed; NVIDIA CI is still required.

---

CI States

Latest PR Test (Base): ✅ Run #33052289112
Latest PR Test (Extra): ❌ Run #33052288823
Latest PR Test (AMD ROCm 7.2): ❌ Run #33052289142

[mickqian]

Reproduction validation

Using the cached Qwen3-VL-2B-Instruct model on an isolated H200 devbox, I injected a one-token multimodal embedding on cache hit (diagnostic-only change, not part of this PR):

• Before this fix: the real serving path logged Got 64 tokens in the text but 1 tokens from multimodal embeddings and terminated the scheduler with Insufficient multimodal embedding length.
• With this fix: the same injected cache corruption was detected and re-encoded; two consecutive multimodal requests both returned HTTP 200, with no embedding-length error.

The injection was applied only in scratch checkouts; no diagnostic code is included in this PR.