feat(embedconfig): expose literal role affixes and dimension requests

[salmonumbrella]

What changed

Expose literal document/query prefixes and suffixes and an explicit request_dimensions flag in embedconfig.Embedder. Empty affixes and a false flag preserve existing identities and requests.

Why

Text retrieval models can need distinct document and query prompts. Applications using the standard TOML configuration currently have to recreate keys that the low-level client already supports. These options let them use the shared configuration directly.

Usage

dims = 768
document_prefix = "title: none | text: "
query_prefix = "task: search result | query: "
request_dimensions = false

Affixes preserve whitespace. dims always validates the returned width; opt into request_dimensions only when the endpoint supports it. Kit never slices vectors. The executable text example describes the serving requirements; model inference remains untested.

Closes #164.

[roborev-ci]

roborev: Combined Review (072ddfd)

Verdict: No findings at or above medium severity.

---

Reviewers: 2x codex, codex (security) | Synthesis: codex | Total: 1m10s

[salmonumbrella]

@wesm can you unblock workflow execution for this PR, approving the runs if GitHub requests approval or re-running them after checking the gate?

Both current-head runs ended action_required before any jobs ran:

• CI 37598216721
• Benchmark Gate 37598216849

Head is 072ddfdd0cc1cf9739816bf7a68afbce16b1c138, one commit with no workflow changes. Current-head Roborev passed. Our account has read access, and the approval-policy endpoint returns 403, so we cannot verify the specific rule or unblock it ourselves.