fix(copilot): route Claude models through Copilot's native anthropic provider

[larsmoan]

Fixes #494.

Claude models routed through Copilot CLI get zero Anthropic prompt-cache
hits, because ucode always sets COPILOT_PROVIDER_TYPE=openai and points
Copilot at the OpenAI-compatible MLflow gateway — a wire format with no field
for a cache_control breakpoint — even for Claude models. Every other agent
ucode configures (claude, opencode, pi) already gets Claude traffic
routed to Databricks' native /ai-gateway/anthropic path; Copilot did not.

Copilot CLI documents (and its shipped runtime implements) a native
COPILOT_PROVIDER_TYPE=anthropic provider — see issue #494 for the
runtime-string evidence (prompt_cache_request.rs, attemptCacheControl,
ephemeral_5m_input_tokens, etc.) that this path actually inserts
cache_control breakpoints. This PR uses it.

Changes

• databricks.py: build_copilot_base_url (single URL) → build_copilot_base_urls
  (dict with anthropic and openai entries), same pattern as
  build_opencode_base_urls/build_pi_base_urls. build_shared_base_urls()["copilot"]
  is now a dict, matching the existing precedent for opencode/pi.
• agents/copilot.py: render_env_overlay now branches on whether the
  selected model is a Claude model ("claude" in model.lower(), same
  substring check databricks.py already uses elsewhere for this family):
  • Claude models → COPILOT_PROVIDER_TYPE=anthropic,
    COPILOT_PROVIDER_BASE_URL = the native anthropic gateway,
    COPILOT_PROVIDER_API_KEY (Copilot's native-Anthropic auth var, sent as
    x-api-key — same header Claude Code's apiKeyHelper already
    authenticates with against this same endpoint).
  • Everything else (codex/gpt-5) → unchanged: COPILOT_PROVIDER_TYPE=openai,
    MLflow gateway, COPILOT_PROVIDER_BEARER_TOKEN.
  • write_tool_config now clears whichever of
    COPILOT_PROVIDER_API_KEY/COPILOT_PROVIDER_BEARER_TOKEN doesn't apply
    before writing, so switching the default model between families doesn't
    leave a stale auth key in ~/.copilot/ucode.env.
• Tests updated/added in test_agent_copilot.py and test_databricks.py.

Not yet verified

I don't have a live workspace handy to confirm system.ai_gateway.usage
actually shows cache_read_input_tokens > 0 for Copilot traffic after this
change — the auth-header compatibility (x-api-key against
/ai-gateway/anthropic) is inferred from how claude.py already
authenticates against the same endpoint, not directly tested against
Copilot's client. Opening as a draft for that reason — happy to run the e2e
suite (tests/test_e2e.py::...test_launch_copilot_per_model) against a real
workspace before marking ready.

Test plan

• uv run pytest tests/test_agent_copilot.py tests/test_databricks.py -q
• uv run pytest -q (full suite; one pre-existing unrelated failure in
  test_e2e_user_agent.py reproduces identically on main)
• uv run ruff check .
• Confirm real cache hits in system.ai_gateway.usage for a Copilot +
  Claude session against a live workspace