[WIP] feat: Add storage checkpoint index for KV offloads

[albertoperdomo2]

Summary

Adds a write-path-only storage checkpoint index so the Go indexer can track which KV blocks exist on shared storage (FS backend). This is the foundation for storage-aware scoring, pre-fetch signals... and any work related to it.

Python side:

• StorageEventPublisher now requires a block_size parameter, populated from hash_block_size * gpu_blocks_per_file in the FS backend manager. This value is emitted in field [4] of BlockStored events so the Go side can derive stride at runtime without duplicated configuration.
• PVC evictor passes block_size=0 explicitly (it only emits removal events).

Go side:

• New StorageIndex interface (AddCheckpoint, HasCheckpoint, RemoveCheckpoint, Clear, SetStride, Stride) backed by a Cuckoo filter in pkg/kvcache/kvblock/storage_index.go.
• Indexer creates and exposes the StorageIndex when StorageIndexConfig.Enabled is true.
• Pool.processEventBatch() detects storage-tier events via isStorageTier() and routes them to the storage index, bypassing the normal engine-key-to-request-key resolution path. Storage events carry canonical request-key checkpoints directly.
• Stride is learned from the first BlockStored event: event.BlockSize / tokenProcessor.BlockSize(). Thread-safe initialization via storageStrideMu mutex with consistency validation on subsequent events.
• Storage indexing is opt-in: when storageIndex is nil, storage events are silently dropped and never fall through into GPU/CPU processing.
• SHARED_STORAGE pod identifier is excluded from scoring to prevent it from appearing as a routable pod.

Testing

Unit tests covered, now benchmarking for potential regressions (WIP).

Related Issues

Closes #520