anthropic: switch opus 4.6/4.7 token thinking budgets to adaptive

[dgageot]

Anthropic's API rejects thinking.type=enabled (token-based extended thinking) for Claude Opus 4.6 and 4.7. When a user configures a token-based thinking_budget for those models, the request now fails at the provider.

This change makes the Anthropic provider transparently switch a token-based budget to adaptive thinking (default high effort) on Opus 4.6 / 4.7 (including dated variants like claude-opus-4-6-20251101), with a slog.Warn so the substitution is visible. Other models keep the existing token-based path unchanged.

Behavior

• claude-opus-4-6 / claude-opus-4-7 (and *-<date>) + thinking_budget: <tokens> → thinking.type=adaptive, output_config.effort=high.
• Same models + explicit thinking_budget: adaptive (or adaptive/<level>) → preserved as-is.
• Same models + disabled / zero / negative budgets (thinking_budget: 0, thinking_budget: none, negative tokens) → passed through untouched, so thinking stays off. (Important: without this guard, those would have been silently flipped to adaptive thinking.)
• Any other model → unchanged.
• max_tokens auto-adjustment is skipped when the budget will be coerced, so an oversized token budget on Opus 4.6/4.7 no longer errors out or inflates max_tokens.

Tests

• TestRequiresAdaptiveThinking covers prefix matching, dated variants, case/whitespace, and non-matches like claude-opus-4-60, sonnet/haiku models, etc.
• TestCoerceAdaptiveThinking covers the coercion path, the no-op paths (already adaptive, non-affected model), and the disabled/non-positive guards.
• TestApplyThinkingConfig and TestApplyBetaThinkingConfig gain Opus 4.6/4.7 cases verifying both SDK type families end up with OfAdaptive set, the right effort, and thinking_display preserved.
• TestAdjustMaxTokensForThinking gains Opus 4.6/4.7 cases verifying no error / no auto-adjust.

golangci-lint run ./... → 0 issues. go test ./... → all packages pass.