Fix prompt-cache round-trip support for ArraysCache, MambaCache, and CacheList

[ronaldmannak]

Proposed changes

This change keeps the existing public savePromptCache / loadPromptCache API intact and extends cache reconstruction so composite and array-backed caches can be restored correctly from persisted prompt caches.

What changed

• ArraysCache now serializes enough metadata in metaState to preserve:
  • total slot count
  • sparse slot positions
  • leftPadding
• MambaCache now round-trips through the same array-cache restore path while preserving its concrete type
• CacheList now serializes child cache structure in metaState and reconstructs children recursively on load
• prompt-cache save/load continues to use the existing safetensors metadata format and Python-compatible cache class names

Tests

Added round-trip coverage for:

• sparse ArraysCache
• ArraysCache with leftPadding
• MambaCache type preservation
• CacheList with KV-only children
• CacheList with hybrid children (MambaCache + KVCacheSimple)

Notes

• No public API changes
• No changes to LRUPromptCache, direct-generation integration, or batching behavior in Add continuous batching support for concurrent LLM inference #150
• This PR is focused on persisted prompt-cache correctness only

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[davidkoski]

Is this OK that it is a NOP? I don't see any calls to set in the code, but other KVCache types do allow set. Should this be a fatalError() instead -- is it a bug to call the setter on CacheList?

[ronaldmannak]

@davidkoski The set is required to conform to the KVCache protocol. The same NOP pattern exists in CacheList.metaState (line 1220) as well, btw. I agree a silent NOP could bite us in the future and I've replaced both with assertionFailure() so they trigger in debug builds but are stripped in release. Happy to switch to fatalError() if you'd prefer a hard crash in all builds.

[davidkoski]

This looks good, thank you. See what you think of my question -- not a blocker for merge.

[davidkoski]

Has a few conflicts from #158 now

[ronaldmannak]

@davidkoski done. should be good to go