feat: add MiniMax M3 provider

[octo-patch]

Summary

Add MiniMax as a new LLM provider, compatible with the existing provider architecture.

Changes

• Register minimax/MiniMax-M3 in MODEL_REGISTRY (1M context window, vision-capable) in nadirclaw/routing.py
• Add minimax / minimax-m3 short aliases to MODEL_ALIASES
• Map the minimax/ model prefix to the minimax provider and the MINIMAX_API_KEY env var in nadirclaw/credentials.py, so credential resolution and provider detection work like the other providers
• Requests route through LiteLLM's native minimax/ provider — no bespoke client
• Add unit tests for alias resolution, provider detection, credential env fallback, cost estimation, and vision capability

Testing

• pytest tests/test_routing.py tests/test_credentials.py — new and existing tests passing

[doramirdor]

Reviewed and verified ✅

• Registration follows the established provider pattern exactly (mirrors the DeepSeek entries): MODEL_REGISTRY, MODEL_ALIASES, PROVIDER_ENV map, and provider-prefix detection map all updated consistently.
• Additive-only (22 lines, no deletions) — no risk to existing models/routing.
• Cost math checks out: 1M in × $0.60 + 1M out × $2.40 = $3.00, matching test_minimax_m3_cost.
• Routes through LiteLLM's native minimax/ provider, no bespoke client needed.
• Ran tests/test_routing.py + tests/test_credentials.py locally → 131 passed. Approved the held CI workflow; it's now green across Python 3.10/3.11/3.12.

Merging.