feat(server): instruction + chat-template wrappers for embedding inputs

[marksverdhei]

Problem

Instruction-tuned embedding models (e.g. LCO-Embedding-Omni / Qwen2.5-Omni) require their inputs wrapped in a task instruction + the chat template, then last-token pooled. The /embedding path tokenizes raw content, so the trained embedding position is never hit and retrieval — especially cross-modal text↔image — collapses to near-random. (This is the follow-up to #126, which stopped images being dropped but left retrieval quality poor.)

Measured on LCO-Embedding-Omni-3B-2605 (last-token, normalize): raw input → cross-modal margin +0.002 (≈random); the correct recipe → +0.30, correct ranking.

Change

Two opt-in, per-modality wrapper flags (empty = off = current raw behavior):

• --embd-prompt-text "…{content}…" — text inputs; {content} → request text.
• --embd-prompt-image "…{media}…" — image inputs; {media} → the server's media marker (one per image), request content ignored.

{media} substitution solves the random-per-restart marker problem, so clients send raw content unchanged (zero client change). \n/\t escapes are processed (string_process_escapes) so values survive INI/preset quoting. Modality is inferred from the request (image_data field / OAI image_url parts). The recipe is symmetric (query == document) so no role field is needed.

Validation

Built + run with the flags via a preset-style literal-\n CLI (proves escape resolution). Sending raw requests, the server applies the recipe:

• text same−cross margin: +0.21 (raw) → +0.30
• cross-modal q="red circle": red 0.65 vs blue 0.35, +0.30 RIGHT (raw: +0.002)
• output L2-normalized ✓

Deployed via the omniswap preset's per-model block (LCO only; qwen3-embed / embeddinggemma stay raw). Follows #126 / #125 / #127.