fix(sdk): serialize audio and video inputs

[svonava]

Related: superlinked/sie-internal#1958

Summary

• serialize audio and video byte/path inputs consistently across Python SDK encode, score, and extract
• add matching TypeScript SDK audio/video Item types and msgpack serialization across encode, score, and extract
• preserve generic structured extractor data in TypeScript ExtractResult (for document/media adapters)
• make TypeScript extract labels optional and map instruction / outputSchema to the generic instruction / output_schema primitive parameters
• document native audio/video Item fields in gateway and local-server OpenAPI
• repair Pydantic nested-schema references when exporting request models

This adds no task-specific endpoint or model behavior. It aligns both public clients and schemas with the existing generic Item wire contract; model adapters and catalog rollout remain separate.

Verification

• Python SDK media + sync/async client tests: 103 passed
• TypeScript focused client/media/type tests: 105 passed
• TypeScript SDK typecheck: passed
• TypeScript SDK ESM/CJS/declaration build: passed
• TypeScript changed-file Biome format/lint: passed
• server OpenAPI export tests: 6 passed
• Ruff format/check: passed
• ty: no new errors (four pre-existing unused-ignore warnings)
• cargo fmt: passed
• gateway binary built and regenerated OpenAPI; generated audio/video refs validated
• git diff --check: passed

The complete TypeScript suite reached 309 passing tests; its only failure is the existing public-checkout wireContract.test.ts dependency on the absent packages/wire-fixtures/model_state.json. The package-wide Biome command also reports the pre-existing package.json formatting drift and two existing encoding.ts non-null warnings; all files changed here pass Biome.

Review posture

Draft for human review. Do not merge as part of automated course work.