Added new model registry

[kapiljain1989]

Summary

• Adds ModelConfig and ModelRegistry to support model-aware KV-cache indexing, enabling HMA (Hybrid Multi-head Attention) vs simple model distinction
• Introduces AttentionGroupConfig with per-group block size and sliding window support for HMA models like DeepSeek-V3
• Integrates model registry into Indexer and Pool so event processing can use model-specific configuration
• Adds ModelAttentionInfo with precomputed attention metadata for efficient scoring lookups

Example
{ "modelConfigs":[ { "name":"DeepSeek-V3", "isHMA":true, "attentionGroups":[ { "groupId":0, "attentionType":"full", "blockSize":64 }, { "groupId":1, "attentionType":"sliding_window", "blockSize":64, "slidingWindowSize":4096 } ] }, { "name":"Qwen/Qwen3-8B", "isHMA":false } ] }

Default Behavior

When no modelConfigs is provided (or the list is empty), the registry treats all models as non-HMA for memory efficiency: