perf(video): accept GPU textures in HW encoders (Vulkan Video, NVENC, VAAPI)

[staging-devin-ai-integration]

Summary

Enable hardware video encoders to accept VideoData::Gpu directly from the compositor, completing the all-GPU pipeline path: decoder/compositor → encoder with zero CPU pixel touches.

Depends on: #350 (GPU VideoFrame variant)
Parent issue: #167 (web renderer context), but this is an independent pipeline improvement

Per-Encoder Work

Encoder	File	Interop Mechanism	Effort
Vulkan Video	vulkan_video.rs	Same wgpu Vulkan device → native VkImage	M
NVENC	nv_av1.rs	cuImportExternalMemory (CUDA↔Vulkan interop)	L
VAAPI	vaapi_h264.rs, vaapi_av1.rs	Export wgpu texture → DMA-BUF → vaCreateSurfaceFromDmaBuf	L

Recommended order

1. Vulkan Video (most natural fit — same Vulkan device as wgpu, no cross-API interop)
2. VAAPI (DMA-BUF is well-supported on Mesa)
3. NVENC (CUDA↔Vulkan interop is more complex)

Architecture

Current path:
  Compositor (wgpu) → readback_rgba() → CPU Vec<u8> → encoder upload → encode

Proposed path:
  Compositor (wgpu) → VideoData::Gpu(wgpu::Texture)
    → encoder receives texture directly → encode (no CPU touch)

Acceptance Criteria

• At least one encoder (Vulkan Video recommended) accepts VideoData::Gpu input
• Encoder checks for GPU data and uses the texture directly when available
• Fallback: if encoder doesn't support GPU input, auto-readback to CPU via to_cpu()
• Benchmark: end-to-end compositor → encode latency with GPU path vs. CPU path
• Document GPU requirements (Vulkan version, driver support)

Complexity

Effort: L (1–2 weeks per encoder)
Each encoder has different interop requirements. Vulkan Video is simplest since it shares the wgpu Vulkan device.

Notes

• This is the final piece of the all-GPU pipeline: web renderer/decoder → compositor → encoder with zero CPU readback.
• Can be implemented incrementally — start with Vulkan Video, add VAAPI and NVENC later.
• The encoder should gracefully fall back to CPU readback if it receives VideoData::Cpu (backward compatible).