perf: cache the OpenRouter model catalog with a 30-minute TTL

[aryansk]

Problem

list_free_models hits https://openrouter.ai/api/v1/models on every call.
/models and startup both fetch the full catalog each time — a wasted
network round-trip and a slow menu on a flaky connection. Closes #56.

Change

Adds an in-process cache to list_free_models in gcode/models.py:

• Successful responses are cached for 30 minutes (monotonic clock), so
  repeated calls inside the TTL make one HTTP request.
• ttl_seconds=0 is the escape hatch for callers that need fresh data.
• Failed fetches are never cached, so the next call retries; the
  best-effort contract (empty list + actionable error) is unchanged.
• The returned list is a copy, so callers can't mutate the cache.
• clear_model_catalog_cache() is exposed for tests/consumers.

Why this approach

In-process only, no new dependencies, and the default matches the issue's
suggested 10–60 min window. Only successful responses are cached so a
transient outage is never masked by a stale catalog, and the escape hatch
keeps /models honest when the user explicitly asks for a fresh list.

Testing

uv run pytest:  112 passed (5 new: cache reuse within TTL, real-clock
                expiry, ttl_seconds=0 bypass, failure-not-cached, copy
                isolation)
uv run ruff check .:  All checks passed
uv run ruff format --check .:  32 files already formatted
uv run mypy gcode:  Success: no issues found in 12 source files

Acceptance criteria covered: two consecutive calls inside the TTL make one
HTTP request (asserted via mock_get.call_count == 1), and cache
invalidation on expiry is tested with the real clock.

Documentation and release impact

• User-facing documentation updated
• Changelog/release note needed
• Migration or compatibility note needed
• No documentation impact

Review notes

• Known limitations: none.
• Follow-up issue: none.