feat: upgrade MiniMax default model to M3

[octo-patch]

Summary

Upgrade the default MiniMax model from M2.7 to M3 across the LLM picker logic, and add the M3 model entries to intentkit/models/llm.csv.

Changes

• Add MiniMax-M3 row to intentkit/models/llm.csv for the direct minimax provider (1M context, $0.6/$2.4 per M tokens, $0.12 cache read, image input).
• Add minimax/minimax-m3 row for the OpenRouter provider (1M context, $0.3/$1.2 per M tokens, image + video input).
• Update pick_summarize_model, pick_default_model, and pick_long_context_model in intentkit/models/llm_picker.py to use MiniMax-M3 instead of MiniMax-M2.7.
• pick_default_model also references minimax/minimax-m3 (was minimax/minimax-m2.7) for the OpenRouter fallback.
• Retain MiniMax-M2.7 and minimax/minimax-m2.7 entries in llm.csv so existing agents pinned to M2.7 keep working.

Why

MiniMax-M3 is the latest MiniMax flagship: a multimodal foundation model with a 1M-token context window built on MiniMax Sparse Attention (MSA) for substantially faster prefill / decode at long context. It also gains native image input support, which the existing Anthropic-compatible integration in MiniMaxLLM can pass through.

Testing

• ruff format / ruff check clean on the changed files.
• python -m py_compile of the changed Python files succeeds.
• CSV schema validated: M3 rows parse correctly, M2.7 rows preserved for backward compatibility, no older M2.5/M2.1/M2/M1 rows present.
• Picker model lookup verified by static parse: MiniMax-M3 is referenced in all three picker functions where M2.7 was previously used; no stale MiniMax-M2.7 references remain in the picker.