feat: upgrade MiniMax default model to M3

[octo-patch]

Summary

Upgrade the MiniMax KV-cache compression validation target to the new flagship MiniMax-M3 while keeping MiniMax-M2.7 available as a fallback. M3 shares the same MoE attention layout (head_dim=128, num_kv_heads=8, num_layers=62, 204K context, GQA 48/8), so the existing IsoQuant / PlanarQuant / LiteratiQuant test surface covers both checkpoints without architecture changes.

Changes

• turboquant/validate_minimax_m2.py
  • Replace the single MODEL_NAME constant with SUPPORTED_MODELS = ("MiniMax-M3", "MiniMax-M2.7") and DEFAULT_MODEL = "MiniMax-M3".
  • Add --model {MiniMax-M3,MiniMax-M2.7} CLI arg (defaults to M3); thread it through run_model_validation and print_memory_estimate.
  • Update module/function docstrings and CLI help text to mention M3 as the new default.
• tests/test_minimax_m2.py
  • Refresh docstrings/comments to describe the shared M3 / M2.7 dimensions; no architectural constants or test bodies changed (all 7 test classes, 18 test functions intact).
• README.md
  • Rename the "## MiniMax-M2.7 Compatibility" section to "## MiniMax-M3 / M2.7 Compatibility" and document --model MiniMax-M2.7 for legacy validation.

Why

MiniMax-M3 is the new flagship in the MiniMaxAI MiniMax family and supersedes M2.7 as the default reference architecture. Surfacing it as the default validation target keeps the project's compatibility story aligned with the latest checkpoint while preserving the M2.7 path so existing users can keep validating against the previous flagship.

Testing

• python -m py_compile turboquant/validate_minimax_m2.py tests/test_minimax_m2.py — both files parse cleanly.
• AST inspection confirms the test module still exposes 7 classes / 18 test functions and the validate module exposes the expected SUPPORTED_MODELS, DEFAULT_MODEL, HEAD_DIM, etc.
• Synthetic benchmark and unit tests are dimension-driven (head_dim=128) and unchanged in logic, so they continue to validate both M3 and M2.7 KV cache compression behavior.