feat: preserve MiniMax model routing - #192

Reason: Complete the MiniMax text catalog rollout after #191 by adding MiniMax-M3.1-Flash and preserving explicitly selected custom or legacy model IDs at the OpenAI-compatible request boundary.

Changes

• Add MiniMax-M3.1-Flash before MiniMax-M3 in the model catalog and settings picker, with the documented one-million-token context budget.
• Route the actual /api/chat request through a shared request builder so model selection and full message payloads are serialized consistently.
• Preserve unknown saved, custom, and legacy model IDs instead of silently replacing them with gpt-3.5-turbo.
• Keep the existing fresh-install OpenAI default and user-configured endpoint behavior unchanged.

Validation

• git diff --check
• Source-backed mocked HTTP-boundary check covering MiniMax-M3.1-Flash, MiniMax-M3, legacy/custom MiniMax-M2.7, and multimodal message parts.
• pnpm lint (passes with no ESLint warnings or errors)
• pnpm exec tsc --noEmit (blocked by existing generated/dependency issues in Prisma exports, react-loader-spinner, and an unrelated nullable Postgres field; no changed file is reported)

Related: #191