Add NV12 output: GPU shader-based RGBA→NV12 conversion (2.8x speedup on Apple Silicon)

[djj0s3]

Summary

When the negotiated downstream format is NV12, projectm now does the RGBA→NV12 conversion on the GPU via two GLSL passes against its existing RGBA FBO, then ReadPixels each plane straight into the GstVideoFrame.

This eliminates the downstream `videoconvert ABGR→NV12` CPU step that dominated render time on Apple Silicon. End-to-end render measured against the same audio (6 min track, 1080p30, vtenc_h264):

```
Before: 363s audio → 842s render (2.3x realtime)
After: 363s audio → 300s render (0.8x realtime)
```

2.8x speedup. Render now completes in less time than the audio plays.

How it works

Two shader passes after `projectm_opengl_render_frame_fbo`:

1. Y pass — full-resolution into a R8 texture, BT.601 luma coefficients
2. UV pass — half-resolution into a RG8 texture; linear-filtered sampling of the source RGBA texture automatically averages 2x2 blocks for 4:2:0 chroma subsampling

`ReadPixels` then pulls each plane straight into the GstVideoFrame's NV12 plane data with `GL_PACK_ROW_LENGTH` accounting for stride alignment.

Caps

```
video/x-raw, format = { ABGR, NV12 }
```

ABGR path is unchanged for any consumer that doesn't accept NV12. `vtenc_h264` advertises NV12 in its sink caps, so the new path is auto-selected when vtenc is downstream.

Implementation gotchas

• VAO is mandatory under GL 3.2 core (macOS Cocoa). Without one, every draw fails with `GL_INVALID_OPERATION`. We bake the vertex attribute setup into the VAO once and rebind per-pass.
• Shader uses `#version 150 core` — works on the GL 3.2 core profile macOS exposes via Tauri/Cocoa. GLES2 not implemented since this targets the local Mac renderer; production GPU pods use a separate `convert_cuda.sh` pipeline.
• PBO async readback is bypassed in NV12 mode — it was a workaround for the slow CPU conversion, which no longer exists.

Testing

• Standalone gst-launch with `audiotestsrc → projectm format=NV12 → vtenc_h264 → mp4mux` produces valid 4:2:0 yuv420p H.264 output (verified with ffprobe).
• Pixel sample of decoded frame shows reasonable color values (not clipped, not all black/white).
• End-to-end bundled-app render with 6 min audio: visually correct output, 2.8x speedup vs ABGR baseline.

🤖 Generated with Claude Code

[revmischa]

please don't commit every preset here

[djj0s3]

Oh sorry! I thought I had forked this work out. Oof, I'll fix. Sorry for the unnecessary noise.

[djj0s3]

Closing — this PR was opened against the wrong repo by mistake. The change lives in our private fork (djj0s3/gst-projectm) and isn't intended for upstream. Apologies for the noise.