[Feature] Optimize session cache performance

[b-re-w]

Issue: No session‑level KV cache leads to redundant re-processing of the full chat history on every request, degrading performance

Implement session-scoped KV cache support

• Persist cache state per session:
  Add attributes (e.g., kv_cache, past_key_values) to Session in api/src/main/settings.py to hold per-session model state. Provide methods to retrieve, update, and clear these caches when a session ends.
• Integrate cache with BaseModel/runtime calls:
  Extend BaseModel.chat and CoreRuntime.__call__ to accept optional cache objects from the Session and return updated cache data after generation.
• Clear cache on session close:
  Modify Session.close to explicitly discard any stored cache state and free related resources.

Add KV cache utilization to llama_cpp backend

• Store and reuse llama.cpp state:
  In api/src/main/backend/gguf.py’s GGUFRuntime.__call__, leverage save_session_file/load_session_file (or similar llama_cpp APIs) to persist key/value caches between calls within the same session.
• Expose cache hooks:
  Accept an incoming cache handle and return the updated handle so Session can manage it. Ensure caches are reset when Session.clean_up or Session.close is invoked.

Enable past key/value caching for Transformers backend

• Manage past_key_values:
  In api/src/main/backend/bin.py, maintain past_key_values from Hugging Face model outputs. Feed this cache back on subsequent calls to avoid re-tokenizing the entire history.
• Session integration:
  Pass past_key_values to/from the Session cache helpers. Reset them on session termination or when the conversation is cleared.