feat: trace the remaining request parameters pi puts on the wire

[milanagm]

Stacked on #29, base is feat/model-parameters, so this diff is only the additions.

Change

This PR adds the rest of what pi actually puts on the wire, each read off the request payload:

chip	source	why
thinking_budget_tokens	payload	the concrete budget the provider was given
prompt_cache_retention	payload	"24h" (OpenAI) / "1h" (Anthropic) — cost-relevant
service_tier	payload	cost-relevant, openai-responses only
tool_choice	payload	string, or the type of the object form
declared sampling params	payload, keyed by model config	temperature, top_p, top_k, min_p, …

Plus context_window in the generation metadata.

Tests

pnpm typecheck clean. pnpm test 87/87, up from 75 on #29.

End-to-end through the real pi CLI:

• prompt_cache_retention: "24h" with compat.supportsLongCacheRetention in models.json and PI_CACHE_RETENTION=long
• declared sampling parameters, asserting the values the wire carried rather than the configured ones
• context_window on every generation