cli, core: add --no-audio option to skip audio VAE decode and write video-only MP4

[lorenzoperrone]

Currently, h3_generate() always decodes the audio VAE (7 blocks) and muxes a soundtrack, even when only visual output is needed. This is primarily a workflow and output-control feature rather than a major performance optimization: on tested Apple Silicon (M2 Pro) it saves roughly 1–4 seconds, while text encoding, DiT, and the video VAE still dominate generation time.

Changes

• Add disable_audio field to struct h3_params and its default initialization in H3_PARAMS_DEFAULT, validated as zero-or-one like other boolean parameters.
• Add --no-audio CLI flag in main.c and document it in usage().
• Bypass the audio VAE decoding phase in h3_generate() when disable_audio is enabled; waveform remains zeroed, ensuring cleanup and error paths remain safe.
• Select the existing video-only writer h3_ffmpeg_write_rgb24() at mux time instead of h3_ffmpeg_write_av_rgb24_f32().
• Add !audio [on|off] command and Audio: on/off display in interactive !status (h3_cli.c).

Use cases

• Fast visual previews and denoising iteration when audio synthesis is not required.
• Creating clean video-only MP4 files without dummy or silence audio tracks, resulting in smaller file sizes.

Verification

Tested on Apple M2 Pro (macOS):

make -j8
make test
./h3 -d ./MiniMax-H3 -p "A cute red fox in snow" --width 256 --height 256 \
    --frames 22 --steps 2 --no-audio -o outputs/test_silent.mp4
ffprobe outputs/test_silent.mp4

1. make test passes all checks cleanly (ok: 1768 checks, FFmpeg mux tests pass).
2. With --no-audio, the generation skips the audio VAE phase entirely.
3. DiT streaming metrics remain identical to the standard path.
4. ffprobe confirms the output container has only a single H.264 video stream and no audio stream:

Stream #0:0[0x1](und): Video: h264 (High) (avc1 / 0x31637661), yuv420p(progressive), 256x256, 24 fps, 24 tbr, 12288 tbn (default)

5. Interactive REPL session tested: !status displays Audio: on, !audio off toggles it to Audio: off, and !audio toggles state back and forth as expected.