Handling Attention Group id in KV events

[kapiljain1989]

Add attention group tracking to KV-Cache indexer for Hybrid Multi-head Attention (HMA) support. This enables per-group cache hit scoring for models with multiple attention groups
(e.g., full attention + sliding window attention).

Changes

• Core data model: Add StoredGroups []int field to PodEntry to track which attention groups have cached a block
• Event schema: Add GroupIdx field to BlockStoredEvent and BlockRemovedEvent for per-group cache updates
• vLLM adapter: Parse group_idx from vLLM KV events (msgpack fields [9] and [3])
• Index implementations: Update all index backends (InMemory, CostAwareMemory, Redis) to:
  • Use string-based cache keys ("podID@tier") instead of struct keys for efficient in-place updates
  • Merge StoredGroups when adding duplicate entries
  • Remove specific groups on eviction (delete entry only when no groups remain)
  • Store JSON-serialized entries in Redis for group list persistence
• Event processing: Convert single GroupIdx from events to StoredGroups list in index operations

Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com