[ENH] Add run_tools_batch MCP tool for read-only agent workflows

[QuantumByte-01]

Motivation

Agentic workflows usually require multiple read-only calls in sequence (e.g., list estimators -> describe candidates -> list available tags/data).
This creates avoidable round-trips and latency for MCP clients.

Proposal

Add a new MCP tool: run_tools_batch

Scope (MVP)

• Accept a list of operations: {tool: str, arguments: dict}
• Execute only read-only tools in batch:
  • list_estimators
  • describe_estimator
  • get_available_tags
  • list_available_data
• Return per-operation structured results with index + tool + success/error
• Reject unsupported tools with clear validation errors

Out of scope for MVP

• Mutating tools (instantiate_*, fit_predict*, release_*, jobs mutation)
• Parallel execution or DAG semantics

Acceptance criteria

• New tool appears in list_tools with schema
• call_tool routes run_tools_batch requests
• Unit tests cover success path, unknown tool in batch, malformed operation entries
• Output is JSON-serializable and stable for agents

If maintainers agree, I can open a focused PR with tests first.

[SHIVANSH-ux-ys]

Hi team! I completely agree that the round-trips for read-only tools were a major bottleneck for agentic workflows. I've gone ahead and implemented the run_tools_batch MCP tool as proposed. It seamlessly processes the list of operations and safely filters out any mutating tools. You can find the implementation in PR #415! Let me know if you'd like me to adjust any of the validation logic.