fix(model): route chat_optimizer through minimax_chat (not just chat_target)

[jcforever1]

Summary

Two missing pieces in the minimax_chat dispatch chain prevented the optimizer side of any SkillOpt training run from working when optimizer_backend='minimax_chat':

1. minimax_backend.py did not define chat_optimizer — only chat_target / chat_target_messages existed. Any caller using optimizer_backend='minimax_chat' would hit AttributeError or silently fall through to the Azure path.
2. skillopt/model/__init__.py chat_optimizer and chat_optimizer_messages dispatchers checked claude_chat and qwen_chat but NOT minimax_chat, so minimax users would silently fall through to _openai.chat_optimizer (Azure) which fails with "Azure OpenAI endpoint is not configured for optimizer." for any user without AZURE_OPENAI_* env vars.

Symptoms before the fix

• Train log repeatedly shows [skip] no usable patches — skill unchanged and total tokens: N (calls=6) — only the target side ran, optimizer was bypassed
• Final summary reports accept=0 reject=0 skip=N regardless of baseline quality, even when baseline is 0% and edits would clearly help
• Affects any user running minimax_chat (or any non-OpenAI/Azure backend that wraps claude-agent-sdk-style chat) as the optimizer — silently produces no training improvement

What this PR changes

• Adds chat_optimizer function to minimax_backend.py (mirrors chat_target via _chat_messages_impl)
• Adds minimax_chat branches to both chat_optimizer and chat_optimizer_messages dispatchers in skillopt/model/__init__.py

Verification

Local 1-epoch training on a 4-item SearchQA-format dataset went from [skip] no usable patches — skill unchanged (baseline 0.5, optimizer bypassed) to accept_new_best with success_patches=1 per step and the skill expanding from 451 chars to 951 chars. Total LLM calls went from 6 (target-only) to 12 (target + optimizer), confirming both sides of the pipeline now run.

Motivation

The skillopt 0.2.0 release notes list "Multi-backend support (OpenAI / Azure / Claude / Qwen / MiniMax)" — but on MiniMax the optimizer side was never actually wired. Anyone who picked minimax_chat based on that list would hit the bug. This was confirmed during a parallel-training bench run on macOS with ANTHROPIC_AUTH_TOKEN routed to MiniMax's anthropic-compat proxy.

Backwards compatibility

• Pure addition (new function + 2 new dispatcher branches); no existing behavior changes
• Existing claude_chat / qwen_chat / openai_chat callers unaffected

Test plan

• Verify skillopt-train runs end-to-end with optimizer_backend=minimax_chat
• Verify skillopt-sleep run works in the same backend (uses chat_target, already worked)
• Run existing searchqa smoke test and confirm success_patches > 0 instead of [skip] no usable patches