feat(sync): add Hugging Face provider sync

[hanouticelina]

Disclaimer: This PR was implemented by Claude Code and tested and reviewed locally by me.

Summary

• add Hugging Face to the shared SyncProvider registry and the hourly sync matrix.
• sync model pricing, context, modalities, and advertised capabilities from the HF Inference Providers /v1/models, resolving each model through base_model metadata in /models
• Hugging Face Inference Providers is an aggregator (many providers per model, routed to the fastest), so pricing and context are taken from the highest-throughput provider, and tool/structured-output support from any provider
• create-only for now: existing curated TOMLs are left untouched.
• add huggingface:sync.

base_model behavior

Each router model is matched to a canonical model already present in /models (via the shared resolveCanonicalBaseModel/factorBaseModel). When a match is found, the generated .toml contains only the fields unique to the router (cost, context, capabilities) and inherits the rest; models with no canonical match or no priced provider are skipped and reported in a notice. HF_TOKEN is optional - the model list is public.

Verification

• unauthenticated GET https://router.huggingface.co/v1/models returned 126 models
• dry run: Dry run: 19 created, 0 updated, 3 removed, 23 unchanged (2
  embeddings + MiMo-V2-Flash) deleted
• bun validate passed and the core typecheck (tsc --noEmit -p packages/core) is clean

  $ bun models:sync huggingface --dry-run

Syncing Hugging Face...
Would create zai-org/GLM-5.2.toml
Would create MiniMaxAI/MiniMax-M3.toml
Would create moonshotai/Kimi-K2.7-Code.toml
Would create Qwen/Qwen3.6-35B-A3B.toml
...
Would remove XiaomiMiMo/MiMo-V2-Flash.toml
Would remove Qwen/Qwen3-Embedding-4B.toml
Would remove Qwen/Qwen3-Embedding-8B.toml
Dry run: 19 created, 0 updated, 3 removed, 23 unchanged

Sync summary
Hugging Face: 19 created, 0 updated, 3 deleted

A created entry stays minimal - the rest is inherited from its base_model:

base_model = "zhipuai/glm-5.2"
open_weights = true

[cost]
input = 1.4
output = 4.4

[limit]
context = 262_144