fix(minimax): honour OPTIMIZER_DEPLOYMENT for optimizer-role calls

Follow-up to #116. MiniMax exposed only TARGET_DEPLOYMENT, and
set_optimizer_deployment() never touched MiniMax, so a configured
optimizer_model was ignored: chat_optimizer and the optimizer message path
both sent TARGET_DEPLOYMENT. In mixed optimizer/target setups this could send
the wrong model name to the optimizer endpoint.

Adds OPTIMIZER_DEPLOYMENT and set_optimizer_deployment() to minimax_backend,
wired into the model facade. chat_optimizer and a new chat_optimizer_messages
use it, falling back to TARGET_DEPLOYMENT when unset. Also fixes dict[int]
return annotations to dict[str, int]. Adds routing tests.

Co-Authored-By: Claude <noreply@anthropic.com>

skillopt/model/__init__.py

@@ -215,7 +215,7 @@ def chat_optimizer_messages(
             timeout=timeout,
         )
     if get_optimizer_backend() == "minimax_chat":
-        return _minimax.chat_target_messages(
+        return _minimax.chat_optimizer_messages(
             messages=messages,
             max_completion_tokens=max_completion_tokens,
             retries=retries,
@@ -540,3 +540,4 @@ def set_optimizer_deployment(deployment: str) -> None:
     _openai.set_optimizer_deployment(deployment)
     _claude.set_optimizer_deployment(deployment)
     _qwen.set_optimizer_deployment(deployment)
+    _minimax.set_optimizer_deployment(deployment)