feat: add direct MiniMax provider support

[octo-patch]

Summary

Add MiniMax as a direct LLM provider with native API integration, supporting both streaming and non-streaming requests through MiniMax's OpenAI-compatible API.

Supported Models

• MiniMax-M2.5 - Peak Performance. Ultimate Value. Master the Complex
• MiniMax-M2.5-highspeed - Same performance, faster and more agile

Both models support 204,800 tokens context window.

Changes

• New file: web/src/llm-api/minimax.ts - Direct MiniMax API handler (follows existing Fireworks/SiliconFlow pattern)
• Model config: Add minimax prefix to ALLOWED_MODEL_PREFIXES and MiniMax models to openrouterModels
• Env schema: Add optional MINIMAX_API_KEY environment variable
• Request routing: Route minimax/* models directly to MiniMax API (takes priority over Fireworks routing)
• Error handling: Add MiniMaxError class with proper error propagation
• Billing: Accurate per-token pricing ($0.30/M input, $1.20/M output, $0.03/M cached input)

Implementation Details

• Uses MiniMax's OpenAI-compatible endpoint (https://api.minimax.io/v1)
• Handles MiniMax's temperature constraint: range (0.0, 1.0], defaults to 1.0 (cannot be 0)
• Follows the exact same architecture as existing direct provider handlers (Fireworks, SiliconFlow, CanopyWave)
• MiniMax models are checked before Fireworks in the routing chain to avoid conflicts with Fireworks' existing MiniMax model map

API Documentation

• OpenAI Compatible: https://platform.minimax.io/docs/api-reference/text-openai-api
• Anthropic Compatible: https://platform.minimax.io/docs/api-reference/text-anthropic-api