feat: improve frontend API, Streamlit UI, FAISS store, and memory storage

[JMcCumberIO]

Frontend API (frontend_api_svc.py):

• Add GET /projects/ endpoint to list all projects (newest-first)
• Add DELETE /projects/{id}/ endpoint with guard against deleting running projects
• Add GET /projects/{id}/metrics/ endpoint exposing prompt/completion token counts
  and a rough USD cost estimate per project
• Add WebSocket /projects/{id}/ws endpoint for real-time streaming of status
  changes, captured logs, and agent messages to connected clients
• Introduce TokenUsage model and _add_token_usage() helper; accumulate token
  usage during both mock and real MetaGPT runs
• Structured JSON-line message logging via _write_message(); messages are
  broadcast over WebSocket in real time
• Improve mock-mode simulation: per-agent-step token accounting and async delays
• Health check now includes in-memory project count

Streamlit UI (metagpt_streamlit_ui.py):

• Move all imports to the top of file (Optional/List/Dict were previously at
  the bottom, causing NameErrors on first use)
• Replace all st.experimental_rerun() calls with st.rerun() (un-deprecated API)
• Add sidebar project list with status icons and one-click project switching
• Add one-click project deletion from sidebar
• Expose the new /metrics/ endpoint as a Cost & Usage Metrics section
• Expose optional project name input on project creation form
• Add helper functions for list_projects and delete_project

FAISS store (faiss_store.py):

• Fix FIXME: call self.persist() after insert_nodes() so added documents
  survive across sessions
• Accept optional metadatas argument and propagate it to TextNode
• Return list of inserted node IDs instead of an empty list

Memory storage (memory_storage.py):

• Replace magic float threshold (0.1) with a configurable similarity_threshold
  parameter (default 0.5) and proper docstring explaining score semantics
• Fix search_similar() to keep results whose score >= threshold (higher score =
  more relevant in llama_index) and respect the k limit
• Guard against None scores from the retriever

https://claude.ai/code/session_01Vz63pW62BgvXDQXJiyeEXp

[JMcCumberIO]

@copilot review this

[Copilot]

@JMcCumberIO I've opened a new pull request, #5, to work on those changes. Once the pull request is ready, I'll request review from you.