feat: Add PVC evictor BlockRemoved events

[albertoperdomo2]

Summary

This PR adds BlockRemoved event emission from the PVC evictor's deleter process, complementing the BlockStored events added in #571. When the evictor deletes KV cache files from shared storage, it now publishes BlockRemoved events so downstream consumers (e.g. the Go storage indexer) can remove stale checkpoints.

The implementation extends StorageEventPublisher with a publish_blocks_removed() method that supports per-call model_name overrides, since the evictor handles files from multiple models on the same PVC. The model_name constructor parameter is optional to just support this multi-model use case.

Block hashes are extracted by reversing the FileMapper filename convention (16-char hex basename), and the original model name is recovered from config.json at the FileMapper base directory. Model name lookups are cached per base directory so only one filesystem read occurs per model per process lifetime. Events are grouped by model and published after each successful batch deletion.

Testing

All existing tests pass. This PR adds the new deleter tests and extends storage event tests.

Validated with:

pytest kv_connectors/pvc_evictor/tests/test_deleter.py -q
pytest kv_connectors/llmd_fs_backend/tests/test_storage_events.py -q

Related Issues

• Add storage KV-cache indexing support #520

Note

Test layout depends on wether #578 is merged before this one or not.

[albertoperdomo2]

cc: @kfirtoledo

[guygir]

Thanks @albertoperdomo2 , overall looks good - just a few small things that need addressing.

[guygir]

When dry_run=True, delete_batch() still returns len(file_paths) as deleted, so this block can publish BlockRemoved even though no files were removed. could we skip event publishing when dry_run is true (and/or only publish after a real delete)?

[albertoperdomo2]

I do not think we should publish BlockRemoved events when dry_run=True. That can cause the Go indexer to drop valid checkpoints (in the upcoming storage indexer), which is data corruption from the indexer's perspective. The event consumer shouldn't need to know about dry runs.