feat(models): add high-priority provider models

[Kamilbenkirane]

Summary

Adds 11 models from provider docs with aligned Celeste constraints (tokens, tools, thinking, multimodal inputs).

Removes duplicate grok-4.20-beta-0309-* catalog entries (xAI API aliases only; use grok-4.20-0309-* in Celeste).

Note: BytePlus Seedance 2.0 requires a ModelArk resource pack on the account before API calls succeed.

Confirmed catalog specs

Provider	Model ID	Ops	max_tokens	Tools / thinking / media	Notes
Cohere	command-a-plus-05-2026	generate, analyze	65 536	thinking_budget −1…65536; images; schema	Command A+ — 128k ctx, 64k out, vision + thinking
BytePlus	seedance-2-0-260128	video generate	—	duration 4–15s; 480p/720p; ref images (≤9), first/last frame	Seedance 2.0
BytePlus	seedance-2-0-fast-260128	video generate	—	same as 2.0	Fast tier
xAI	grok-4.3	generate, analyze	131 072	WebSearch, XSearch, CodeExecution; images	grok-4.3 — 1M ctx
xAI	grok-4.20-0309-reasoning	generate, analyze	131 072	same tools; images	docs
xAI	grok-4.20-0309-non-reasoning	generate, analyze	131 072	same tools; images	docs
xAI	grok-imagine-image-quality	image generate, edit	—	same aspect ratios as grok-imagine-image	image-quality
OpenAI	gpt-5.5	generate, analyze	128 000	thinking minimal→xhigh; verbosity; WebSearch; images/docs	GPT-5.5
OpenAI	gpt-5.4-mini	generate, analyze	128 000	thinking minimal→high; WebSearch; images/docs	400k ctx per OpenAI
OpenAI	gpt-5.4-nano	generate, analyze	128 000	same as mini tier	Cost-efficient tier
Mistral	mistral-large-2512	generate, analyze	32 768	tools (generic); images/docs; schema	Large 3 — 256k ctx
Mistral	mistral-small-2603	generate, analyze	32 768	tools; images/docs; schema	Small 4 — 256k ctx

Already in catalog (no change): gemini-3.1-flash-tts-preview (audio).

Test plan

• make ci passed locally

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.