Accept AI Gateway V3 during preflight

[david-siqi-liu]

Summary

• probe AI Gateway V3 first through the bounded Unity Catalog model-services endpoint, then fall back to V2
• pass preflight when either version is available while preserving definitive invalid-token short-circuiting
• treat endpoint-specific 403 responses as ambiguous, try the alternate version, and report tailored V3 grant or V2 workspace-permission guidance if both probes fail
• use the version-agnostic preflight in configure/launch flows and e2e coverage

Testing

• UV_CACHE_DIR=/tmp/ucode-uv-cache uv run pytest tests/test_databricks.py -k EnsureAiGateway -v (14 passed)
• UV_CACHE_DIR=/tmp/ucode-uv-cache uv run pytest tests/test_cli.py -k ConfigureSharedState -q (20 passed)
• UV_CACHE_DIR=/tmp/ucode-uv-cache uv run ruff check .
• earlier full-suite run: 1846 passed, 39 skipped; one unrelated existing MCP client-order assertion failed (tests/test_mcp.py::TestConfigureMcpCommand::test_registers_discovered_external_server)