fix: embedding driver honors base_url override and embedding_base_url config (fixes #1212)

[BunnyMoth]

Summary

Fixes #1212. The embedding driver hard-coded 6 cloud providers and
ignored all base_url override paths, forcing self-hosted deployments
to leak content to OpenAI cloud even when a local embedding server
was configured.

Root Cause

The embedding driver cached its URL at init and never consulted
[providers.openai].base_url or [provider_urls]. The needs_v1
allowlist also prevented non-listed providers from appending the
correct /v1/embeddings path suffix.

Changes

Three-tier URL resolution chain (kernel.rs)

All embedding driver init sites now resolve the URL through a
consistent chain:

1. config.memory.embedding_base_url — new explicit embedding override
2. config.provider_urls.get(provider) — existing per-provider override
3. config.default_model.base_url when embedding provider matches
   the LLM provider — the missing link that caused the regression

New embedding_base_url field in MemoryConfig (config.rs)

[memory]
embedding_provider = "openai"
embedding_model = "nomic-embed-text-v1.5"
embedding_base_url = "http://my-llama-cpp:9080/v1"

Removed needs_v1 provider allowlist (embedding.rs)

Replaced the hard-coded allowlist with a simple trailing-slash trim.
The driver now trusts the URL the caller supplies, exactly like
OpenAIDriver does. Default URLs already carry the correct path suffix.

Testing

• cargo fmt --all --check — clean
• cargo clippy --workspace --all-targets -- -D warnings — zero warnings
• cargo test --workspace — 997 tests, 0 failures
• New tests: test_create_embedding_driver_openai_custom_base_url,
  test_create_embedding_driver_openai_custom_base_url_no_v1,
  test_create_embedding_driver_cohere_custom_url,
  test_create_embedding_driver_unknown_provider_custom_url