feat(minimax): add MiniMax-M3 to minimax and minimax-cn providers

[kapelame]

What

Adds the new flagship MiniMax-M3 model entry to both minimax (international) and minimax-cn (China) providers.

M3 spec

field	value
context	512_000
output	196_608 (server hard cap)
attachment / multimodal	true — text + image input
reasoning	true (Anthropic-style thinking blocks)
tool_call	true
API	Anthropic Messages + OpenAI chat compatible (same endpoints as M2.7)

Why this matters for downstream consumers

Tools that read this catalog (Hermes Agent, OpenClaw, opencode, kilo, vercel-ai-sdk wrappers, etc.) currently default to M2.x context windows and text-only modalities for any unknown minimax model. Without this entry, M3 users get silent image stripping and 204K-cap behavior in those tools.

Files

providers/minimax/models/MiniMax-M3.toml      (new)
providers/minimax-cn/models/MiniMax-M3.toml   (new)

Verified

• bun ./packages/core/script/validate.ts exits 0 with both new entries serialized correctly (attachment: true, modalities.input: ["text","image"], limit.context: 512000, limit.output: 196608).

Notes

• Pricing fields placeholder = M2.7 prices. Will update once final M3 pricing is published.
• This PR only adds M3. M2.x attachment flag and other historical entries are intentionally untouched.

[kapelame]

Closing temporarily for internal review. Will reopen once verified.

[kapelame]

Superseded by #1929 (rebased with finalized M3 spec: 512K context, correct pricing, multimodal). This PR's head commit was rewritten so it can't be reopened.