fix: support music covers and surface stream errors

[NianJiuZst]

What

• accept the official music-cover and music-cover-free models in sdk.music.generate()
• validate cover-specific prompt, lyrics, and reference-audio constraints separately from text-to-music generation
• support both one-step covers (audio_url / audio_base64) and two-step covers (cover_feature_id + replacement lyrics)
• make the shared audio stream decoder propagate API/business errors and malformed payloads instead of returning an empty successful stream
• preserve audio carried by the terminal status: 2 stream event and validate streamed hex data
• document SDK cover usage and add regression coverage

Why

This is a follow-up to #196. The CLI gained the current music models and cover workflow, but the SDK still validated every request as text-to-music generation. As a result, it rejected both official cover models and incorrectly required lyrics for one-step covers.

The shared streaming decoder also skipped JSON parse failures, API error payloads, and terminal frames. A 200 application/json response containing a non-zero base_resp.status_code could therefore finish as a zero-byte “success”.

Impact

• SDK users can call both official cover models through the existing music generation endpoint.
• Invalid cover combinations fail locally with actionable usage errors.
• Music CLI and SDK streaming paths now fail closed on API errors, malformed chunks, invalid hex, and streams with no audio.
• A non-SSE JSON success containing hex audio is decoded as a safe fallback.

Validation

• bun test — 400 pass, 0 fail
• bun run typecheck
• bun run build:dev
• bun run lint — 0 errors; 4 pre-existing warnings outside the changed files
• targeted ESLint on changed TypeScript files — clean
• git diff --check