fix(model): complete MiniMax optimizer role deployment, timeout forwarding, and typing

[RohithPariki]

Problem

In skillopt.model.minimax_backend and the top-level skillopt.model dispatcher:

1. chat_target dropped the timeout parameter when routing to minimax_chat, ignoring caller-configured timeouts.
2. chat_optimizer_messages was not implemented in minimax_backend.py, forcing the dispatcher to fall back to _minimax.chat_target_messages(..., stage="optimizer") with mismatched target deployment resolution.
3. set_optimizer_deployment was missing in minimax_backend.py, causing skillopt.model.set_optimizer_deployment() to omit MiniMax from optimizer deployment updates.
4. Return type annotations in minimax_backend.py (dict[int]) were malformed generics.

Root Cause

MiniMax backend support previously focused on the target path, leaving optimizer role deployment registration, timeout forwarding, and structured message dispatcher entry points incomplete.

Solution

• Added OPTIMIZER_DEPLOYMENT and set_optimizer_deployment to skillopt.model.minimax_backend.
• Added chat_optimizer_messages in minimax_backend.py using OPTIMIZER_DEPLOYMENT.
• Forwarded timeout across chat_target, chat_optimizer, and chat_optimizer_messages.
• Fixed return type annotations to tuple[str, dict[str, int]].
• Registered _minimax.set_optimizer_deployment in skillopt.model.set_optimizer_deployment.
• Added unit tests in tests/test_minimax_backend.py validating deployment resolution, tool calls, and timeout propagation.

Testing

• Ran unit tests: pytest tests/test_minimax_backend.py (9/9 passed).
• Ran lint check: ruff check (clean).

Risk

Low. Additive and backwards-compatible with existing backend configuration.

[Yifan Yang (Yif-Yang)]

Thanks — the role-specific MiniMax routing and timeout forwarding are useful. One initialization blocker remains: chat_optimizer_messages() reads OPTIMIZER_DEPLOYMENT, but the module never initializes that global; it is created only if set_optimizer_deployment() has already been called. A fresh process calling the optimizer path therefore raises NameError. Please initialize it from the optimizer deployment environment/default at import time and add a fresh-import regression that calls the optimizer directly without invoking the setter first.