Extract walk_catalog_schemas scaffold in databricks.py

[xsh310]

What changed, and why?

Change: Extract the catalogs → schemas → parallel-per-schema-probe walk out of list_all_mcp_services into a generic walk_catalog_schemas[T](*, deadline, probe, collect, skip_catalogs=…) in databricks.py, and rewrite list_all_mcp_services on top of it. The scaffold owns catalog/schema enumeration, the worker pools, and the wall-clock deadline drain; the caller supplies the probe (list_mcp_services) and a collect that owns accumulation, dedup, progress, and streaming.

Why: Pure refactor (PR B in the ug skill add interactive-picker stack). Extracting the walk lets PR D's skills discovery reuse it as a second caller. list_all_mcp_services's signature and every return value are unchanged.

Note: the design doc's PR B also mentioned rewriting list_uc_functions_catalog_schemas, but that function does not exist in this repo — there is a single walk today, so the scaffold has one caller now and the skills walk becomes the second in PR D.

How do you know it works?

Testing: ruff check and ruff format clean; full uv run pytest green except the two pre-existing e2e failures (test_e2e_user_agent, test_claude_smart_routing_v2) that also fail on main. Added TestWalkCatalogSchemas (skips skip-catalogs + information_schema, probes each user schema, reports progress, returns the right phase-1 reason); the existing TestListAllMcpServices tests stay green unchanged.

This pull request and its description were written by Isaac.