feat: upgrade MiniMax models to M3

[octo-patch]

This PR upgrades the MiniMax models listed in this curated list to the new M3 default.

Changes

• Kilo Code: Replace minimax/minimax-m2.5:free (196K context, 8K output) with minimax/minimax-m3:free (512K context, 128K max output, supports image input)
• NVIDIA NIM: Add new minimax/minimax-m3 (512K context, 128K max output, supports image input) alongside the existing minimax/minimax-m2.7
• OpenRouter: Replace minimax/minimax-m2.5:free (196K context, 8K output) with minimax/minimax-m3:free (512K context, 128K max output, supports image input)
• Kilo Code description and footnote 5 updated to reference M3
• README.md regenerated from data.json via node scripts/generate-readme.js

M3 vs M2.x

	M3 (new default)	M2.5 (removed)	M2.7 (retained on NVIDIA NIM)
Context	512K	196K	128K
Max Output	128K	8K	8K
Modality	Text + Image	Text	Text

M3 is the new default, M2.7 is preserved as a legacy option, and M2.5 (older than M2.7) is removed.

Files changed

• data.json (source of truth)
• README.md (regenerated)

data.json is regenerated to README.md automatically on merge by .github/workflows/generate-readme.yml.

[SebConejo]

Thanks. Before this can move: the OpenRouter page for minimax-m3:free lists a 1M context and image plus video input, not 512K and Text + Image. Can you reconcile that and link sources for the Kilo Code and NVIDIA NIM figures? I verify entries with a real request before merging, so the exact ids and limits matter.