fix(antigravity): read the Gemini pool from the quota summary reliably

[fix2it]

Summary

Fixes Antigravity quota retrieval so that the first-party Gemini quota pool is correctly identified, parsed, and displayed without falling back to or mixing with third-party model groups.

Motivation & Problem

1. User-Agent Filtering: The Google Cloud Code internal quota endpoint (daily-cloudcode-pa.googleapis.com) inspects the client User-Agent. When queried with an unqualified antigravity header, the backend filtered out the Gemini Models group entirely, returning only third-party Claude/GPT groups.
2. Ambiguous Group Selection: When the Gemini group was not explicitly matched, the poller picked the first available group with windows, causing third-party (Claude/GPT) usage percentages and reset times to be displayed under the Antigravity (Gemini) widget.
3. Artificial Day Shifts: The poller previously applied a speculative $+6\text{ day}$ offset to weekly buckets, resulting in inaccurate countdown horizons.

Key Changes

1. Explicit Client Version: Send User-Agent: antigravity/2.24.0 across loadCodeAssist, fetchAvailableModels, and retrieveUserQuotaSummary. This ensures Google's API returns the complete, modern quota tree including Gemini models.
2. Authoritative Summary Endpoint:
   • The first endpoint returning a valid retrieveUserQuotaSummary is treated as authoritative and returned immediately.
   • Endpoints are no longer inconsistently merged, and secondary fetchAvailableModels calls are skipped when the summary has answered.
3. Strict Group Classification:
   • Groups are strictly classified as Gemini vs. External (claude, gpt, 3p-* buckets).
   • An unrecognized group is used only if it is the sole group present.
   • A summary with empty groups or unrecognized windows is treated as "no limit data" rather than an error.
4. Clean Reset Parsing:
   • Uses direct ISO 8601 parsing (parse_iso8601) on bucket reset timestamps without artificial day offsets.
5. Accurate Serialization:
   • Unreported windows are cached with percentage: null (displaying -- rather than a misleading 0% used).

Verification & Tests

• Added unit tests in src/poller/tests.rs covering summary parsing with multiple groups, single-group fallbacks, missing windows, and empty buckets.
• Verified against live Google Cloud Code pa endpoints with real credentials (live_quota_summary_shape).