Standardize Block Hashing for KVCache Compatibility

[sagiahrac]

To ensure consistent hashing from prompt tokens to block hashes that align with the underlying KVCache index, the block hasher logic must be factored out into a single, shared implementation.

Right now, if each prefix store implements its own block hashing, there’s a risk of divergence. Instead, we should rely on one canonical block hasher (compatible with vLLM’s prefix caching design) so that block boundaries and hash values are uniform across all stores.

[vMaroon]

There would be no problem if prefix-stores implemented different hashing - as this is purely a caching optimization to avoid excessive tokenization. The place things can fall is the TokenProcessor hashing - as that has to be aligned with vLLM at the moment (and will continue to be unless we see a strong reason for decoupling, which comes at costs such as carrying token-ids in KV-events). Since the latter expands at vLLM's pace in the domain, it is rather static - but minimal interfacing can increase readability and extensibility.