Add opper-multimodal skill (media generation + realtime)

[Jose-Sabater]

Why

We've shipped a lot of multimodal + realtime surface recently (images, audio TTS/STT, video, OCR, /v3/files, realtime voice). It was crammed into opper-api, whose frontmatter description was near the 1024-char cap and whose body mixed text/gateway concerns with media. A user asking "generate a video" or "transcribe audio" deserves a skill whose description triggers precisely.

What

New standalone sub-skill opper-multimodal (sibling of opper-api/opper-cli/opper-sdks), covering everything beyond text:

• Media generation: POST /v3/images, /v3/audio/{speech,transcriptions}, /v3/videos (async), /v3/ocr — with scoped discovery lists, the shared media contract (store/file_id/parameters passthrough), and curl seeds.
• /v3/files storage API + file_id reuse.
• Vision / PDF input on chat models (rides compat, cross-linked).
• Realtime voice: wss://api.opper.ai/v3/realtime, browser tickets via POST /v3/realtime-sessions, function-scoped /v3/realtime/{name}, providers, caps, billing.

Scope boundary: server-side tools (opper:web_search etc.) intentionally stay in opper-api — they ride the compat chat endpoints.

Wiring (everything stays connected)

• opper-api: collapsed the heavy multimodal/realtime sections into a short pointer to opper-multimodal; kept server-side web search; tightened frontmatter + gotchas + "where to look next".
• opper router: added the new skill to the decision table, Phase-2 proposals, Phase-3 fetch block (primary + GitHub fallback), Phase-4 test rows, vocabulary surfaces, and the sub-skill index.
• README: new skills-table row, example prompts, install lines, "all five skills".
• Every surface points outward to the live OpenAPI spec (https://api.opper.ai/v3/openapi.yaml) and docs.opper.ai — endpoint shapes were verified against the live spec; the spec/docs themselves are untouched.

scripts/validate-skills.sh and scripts/build-site.sh --dry-run both pass (5 skills; new file 205 lines, description under cap).