docs: verified example — embed via any OpenAI-compatible API - #6

XERJ can embed `semantic_text` through an external OpenAI-compatible
`/v1/embeddings` endpoint (`--embed-mode proxy`), but there was no runnable
proof of it. This adds a key-free, reproducible one.

`docs/examples/external-embeddings/`:
  * mock_embed_server.py — a ~40-line dependency-free server implementing the
    exact `{input,model} -> {data:[{embedding,index}]}` contract XERJ sends.
  * xerj-proxy.toml — the `[embedding] mode=proxy` config wiring XERJ to it.
  * run.sh — builds nothing, starts the mock, runs XERJ in proxy mode, indexes
    two docs, and shows the external API is called at ingest AND at query time.
  * README.md — how it's wired, the wire contract, a table of real endpoints
    (OpenAI / Gemini OpenAI-compatible / local), and the ingest=query embedder
    symmetry that keeps vectors comparable.

Verified end to end on this branch: XERJ logs `embedding backend: external
proxy`; indexing two docs produces two `/v1/embeddings` calls; a semantic query
for "database outage from too many open connections" is embedded by the same
API (third call) and correctly returns the "connection pool was exhausted"
document by vector similarity.

Honesty note kept in the README: verified against the reproducible mock (which
speaks the exact contract). Real providers using the same OpenAI contract are
drop-in, but endpoints/model names change — the doc points readers to each
provider's current docs rather than freezing a claim.

Docs/example only; no engine code changed.