fix(proxy): clamp xhigh reasoning effort to high for o-series models - #7370

Follow-up to the second suggestion in #7318's review, now extended with the grok-4.3 note from #7369's review.

Problem / 根因

Since #7318 passes xhigh through verbatim (and max / thinking: adaptive already mapped to it before), two model groups receive a reasoning_effort value they do not accept:

• o-series (o1 / o3 / o4-mini) — accepts low / medium / high only; xhigh is invalid.
• grok-4.3 — accepts low / medium / high but not xhigh (per the xAI docs, noted in fix(proxy): accept the whole grok-4.x (x>=5) family in the reasoning-effort whitelist #7369's review); it was also missing from the whitelist entirely, so its reasoning effort was silently dropped.

Fix

1. Whitelist: accept the grok-4.3 family (grok-4.3, grok-4.3-*) in supports_reasoning_effort, alongside the merged grok-4.x (x ≥ 5) rule.
2. Clamp: clamp_reasoning_effort(model, effort) downgrades xhigh to high when the target is an o-series model or grok-4.3 — the closest supported intent, mirroring how grok-4.5 documents a silent xhigh→high fallback. Applied at both injection sites (Chat Completions reasoning_effort, Responses reasoning.effort). Every other effort value and every other model is byte-identical to before; the new max tier (gpt-5.6/6 family, supports_max_reasoning_effort) never passes through the clamp.

Rebased onto current main (includes #7369's rule and the max-tier resolution).

Tests

• Unit: whitelist accepts grok-4.3 / grok-4.3-something; clamp downgrades xhigh for o3-mini / o1 / grok-4.3 (incl. mixed case) and passes through for gpt-5.4 / grok-4.6 / grok-4.7 and non-xhigh efforts.
• Chat path: explicit xhigh and thinking: adaptive against o3-mini emit high; xhigh against grok-4.3 emits high while high passes through.
• Responses path: xhigh against o3-mini emits high; against gpt-5.4 stays xhigh.
• cargo test --lib 2991 passed / 0 failed (one known local port-binding flake skipped, unrelated); cargo fmt --check clean.

Happy to adjust.