fix(google): support regional model prefixes and Gemini 3 Pro minimum thinkingLevel

[zchryst]

Background

1. Regional endpoint prefixes (such as au.gemini-3.5-flash, eu.gemini-3.5-flash, and us.gemini-3.1-pro) were not matched by /(^|\/)gemini-/i in getGoogleModelCapabilities, causing usesGemini3Features to evaluate to false and routing Gemini 3+ models to the Gemini 2.5 thinkingBudget configuration instead of thinkingLevel.
2. getMinimumThinkingLevelForGemini3Model only checked for gemini-X.Y-flash and defaulted to minimal whenever versionMatch == null. Because Gemini 3+ Pro models (gemini-3.5-pro, gemini-3.1-pro, gemini-3.1-pro-preview, gemini-3-pro-preview) do not support thinkingLevel: minimal (minimum supported level is low), passing reasoning: none or minimal resulted in an HTTP 400 INVALID_ARGUMENT error from the Gemini API.
3. Merging resolvedThinking with providerOptions.google.thinkingConfig could leave both thinkingLevel and thinkingBudget defined simultaneously when a Gemini 3 model had top-level reasoning set alongside providerOptions.google.thinkingConfig.thinkingBudget (or vice versa).

Summary

• Updated Gemini capability and version regexes in packages/google/src/google-model-capabilities.ts and packages/google/src/google-language-model.ts from /(^|\/).../i to /(^|[/.]).../i so regional endpoint prefixes (au., eu., us.) are properly recognized.
• Updated getMinimumThinkingLevelForGemini3Model in packages/google/src/google-language-model.ts to strip regional dot prefixes and return low for all Gemini 3+ Pro models (/^gemini-(?:\d+(?:\.\d+)?-)?pro(?:$|-)/).
• Cleared thinkingLevel when providerOptions.google.thinkingConfig.thinkingBudget is explicitly provided (and vice versa) so both fields are never sent simultaneously.
• Added unit tests in google-model-capabilities.test.ts and google-language-model.test.ts covering gemini-3.5-flash, au.gemini-3.5-flash, eu.gemini-3.5-flash, gemini-3.5-pro, gemini-3.1-pro, us.gemini-3.1-pro, and providerOptions.google.thinkingConfig override precedence.
• Added a patch changeset for @ai-sdk/google and @ai-sdk/google-vertex.

End-to-End Verification

Verified generated generationConfig.thinkingConfig payloads across @ai-sdk/google and @ai-sdk/google-vertex for gemini-3.5-flash, au.gemini-3.5-flash, gemini-3.5-pro, gemini-3.1-pro, and gemini-2.5-pro with reasoning: none, minimal, low, medium, high, and providerOptions.google.thinkingConfig overrides.

Checklist

• All commits are signed (PRs with unsigned commits cannot be merged)
• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)