fix(mlx): validate vision prefix cache fail closed

[RezaGolriz]

Summary

• require positive media-region validation before reusing a vision KV prefix
• truncate the reusable prefix before the first unvalidated image region
• reject missing, shifted, duplicated, empty, or hash-mismatched media metadata instead of silently trusting it
• preserve normal text-only prefix-cache behavior

Fixes #2272.

Why this approach

Image placeholder tokens can be identical even when the image bytes differ. A token-prefix match is therefore insufficient once it reaches a media region. The cache now fails closed unless the cached and incoming media spans have a one-to-one positional and hash match.

Validation

• uv run pytest -q src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py -m 'not slow' — 28 passed
• uv run pytest -q src/exo/worker/tests/unittests -m 'not slow' — 161 passed, 184 deselected
• uv run basedpyright src/exo/worker/engines/mlx/cache.py src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py
• uv run ruff check src/exo/worker/engines/mlx/cache.py src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py

The regression tests cover different image hashes, missing metadata, positional shifts, duplicate spans, empty hashes, exact matches, and text-only prefixes.