Pin prediction cache to resolved model revisions

[LarryHu0217]

What does this PR do?

Include the immutable Hub revision already resolved by the Transformers and VLM backends in the prediction cache identity.

This prevents a mutable revision such as main from reusing predictions generated by older model weights. Backends that do not expose an immutable identity retain their existing cache hashes.

The caching documentation now also explains the remaining limitation for local model directories and opaque endpoints whose weights are replaced in place.

Addresses #1364

Testing

• Focused cache revision and Transformers tests: 2 passed, 8 deselected, 3 subtests passed
• Runnable cache unit-test subset: 6 passed, 1 skipped, 3 deselected, 15 subtests passed
• Ruff check: passed
• Ruff format check: passed

The full cache module was also attempted locally. Tests requiring the optional vLLM and torchvision backends could not run because those dependencies are not installed.