feat: custom OpenAI-compatible embedding providers

[long2ice]

What

Extends the custom-provider mechanism (#117/#212) to embeddings, so any OpenAI-compatible /embeddings endpoint (self-hosted vLLM / Ollama / LM Studio, or a remote gateway) can be added and routed — and unifies how custom chat + embedding models are managed.

Why

Custom providers previously only worked for chat models (models table + chat fallback chain). Embeddings had no custom support at all — callProvider's switch(platform) only covered the 7 built-in platforms. This adds parity.

Changes

Server

• embedding_models gains a key_id column binding a custom row to its endpoint's api_keys row (migrateEmbeddingsV2Custom); built-ins stay NULL and resolve by platform as before.
• Embeddings service resolves a custom row's key + base_url and calls the endpoint via the existing OpenAI-style adapter; adds probeEmbeddingDimensions.
• POST /api/embeddings/custom — probes the endpoint to auto-detect the vector dimension, then upserts. Joining an existing family requires a matching dimension (a family is one vector space) and lands at the back of that family's chain.
• DELETE /api/embeddings/custom/:id and DELETE /api/models/custom/:id — remove a single custom model while keeping the shared endpoint key.
• Adding a model to an existing endpoint updates key/label independently and only when supplied, so adding a second model with a blank key no longer clobbers the key the endpoint's other models depend on.

Client

• Keys page: the two separate custom forms are merged into one "Add a custom model" form with a Chat/Embedding toggle. Several chat + embedding models can share one endpoint and key. Embedding mode auto-detects the dimension and suggests a family by model id (overridable; the dimension check still guards).
• Embeddings page and Models (fallback) page: per-row delete for custom providers (built-ins remain non-deletable).

Routing semantics

Embeddings still only fail over within a family (same model, another provider) — vectors from different models are incompatible. A custom endpoint either becomes its own family (default) or joins an existing one for cross-provider redundancy when dimensions match.

Tests

345 passing. New coverage: custom embedding routing + within-family failover, dimension probe + family-join dimension guard, blank-key reuse of an existing endpoint key, and single custom-model deletion (keeps the shared key; 404 on built-ins).