refactor(providers): share embedding transport

[gold-silver-copper]

Summary

• add an explicit OpenAIEmbeddingsCompatible capability with typed provider policy for endpoint selection, supported fields, usage requirements, and dimension spelling
• move Together, OpenRouter, and Mistral embeddings onto the shared generic transport while preserving OpenAI and Llamafile behavior
• preserve raw provider errors, validate response lengths, and handle provider-specific usage requirements
• serialize one typed compatible request instead of mutating serde_json::Value, while keeping each provider's policy beside its embedding model
• reject Base64 before HTTP for every provider using the shared numeric-vector response parser

Fixes #2078

Provider/API notes

• Together uses /v1/embeddings, omits unset dimensions, and rejects inherited encoding_format/user fields.
• OpenRouter uses /api/v1/embeddings and accepts optional usage, dimensions, encoding format, and user fields.
• Mistral uses /v1/embeddings; output_dimension is emitted only for codestral-embed/codestral-embed-2505 and is capped at the documented maximum of 3072. Fixed-size models reject configured dimensions, and user is rejected before HTTP.
• Base64 response encoding is rejected consistently before HTTP by the shared transport because its response parser accepts numeric vectors; this covers OpenAI, OpenRouter, Together, Mistral, and Llamafile.
• OpenAI's public EmbeddingResponse still requires usage, while the private compatible response shape permits providers that omit it.

Verification

• cargo fmt --all -- --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test -p rig-core --lib embedding (54 passed, 2 ignored)
• cargo check -p rig-core --target wasm32-unknown-unknown --features wasm
• cargo test
• cargo doc --workspace --no-deps (passes with one pre-existing rig-neo4j private intra-doc-link warning)
• cargo test -p rig --all-features --test openai openai::cassette -- --nocapture --test-threads=1 (86 passed; replayed, no cassette changes)

Three independent full-diff reviews were performed. Earlier reviews identified incorrect Mistral dimension shaping and Base64 exposure; both were fixed with regression tests. The final typed-request review found no P0/P1/P2 issues.

AI assistance

This implementation was developed with AI assistance. Every changed line and the complete merge-base diff were reviewed, provider fields were checked against official schemas/documentation, and the verification above was run locally.