clear stale MCP entries when switching workspaces

[asujithan]

Summary

Fixes #110. After switching Databricks workspaces in ucode, agent configs (~/.claude.json, ~/.codex/..., etc.) used to keep MCP entries pointing at the previous workspace until the user explicitly re-ran ucode configure mcp. This PR scrubs them automatically.

What Changed

• New helper purge_cross_workspace_mcp_residue(state, workspace) in mcp.py runs two passes:
  1. Drops entries from state["mcp_servers"] whose URL host doesn't match the current workspace and unregisters them from installed clients.
  2. Removes MCP names tracked only in other workspaces' state buckets from installed clients (idempotent — re-runs stay quiet once configs are clean).
• Called from two places:
  • configure_mcp_command — heals state pollution from before this fix existed.
  • configure_shared_state — runs on every workspace switch, so users no longer need a follow-up configure mcp to reconcile agent configs.

Test plan

• uv run ruff check .
• uv run ruff format --check src/ tests/
• uv run pytest --ignore=tests/test_e2e.py
• New focused tests in tests/test_mcp.py cover both cleanup passes and the idempotency case.
• New tests in tests/test_cli.py verify the switch-time trigger fires only when the workspace actually changes.