Add MiniMax-M3 and MiniMax-M2.7 input token limits

[octo-patch]

Reason: parameter-refresh — add MiniMax-M3 (1,000,000) and MiniMax-M2.7 (204,800) input token limits so the OpenAI-compatible client stops falling back to the 131,072 default.

Changes

• packages/core/src/core/tokenLimits.ts: add vendor-declared exact context-window entries (minimax-m3: 1,000,000; minimax-m2.7: 204,800) to LIMITS, and add first-match regex patterns for minimax-m2.7 (more specific, dotted) and minimax-m3 to PATTERNS. Values are taken verbatim from the vendor-declared context windows rather than the power-of-two approximations used elsewhere, matching the existing convention for vendor-declared decimal limits (e.g. 200k -> 200_000).
• packages/core/src/core/tokenLimits.test.ts: add a MiniMax test suite covering case-insensitivity, the MiniMax-M3 and MiniMax-M2.7 limits, date-suffix normalization, and explicit assertions that neither model falls back to DEFAULT_TOKEN_LIMIT.

Checks

• vitest run src/core/tokenLimits.test.ts — 65 tests passed (including the new MiniMax cases).
• prettier --check on both modified files — all matched files use Prettier code style.
• tsc --noEmit in packages/core — no errors introduced by this change (the only reported error is a pre-existing missing @types/uuid declaration in src/utils/openaiLogger.ts, unrelated to this diff).