feat: add Wan2.2 S2V (speech-to-video) support

[noctrex]

Summary

Audio-driven video generation for Wan2.2-S2V-14B: a reference portrait plus a driving WAV produce a lip-synced video whose output container carries the driving audio.

• detect S2V checkpoints by audio-module key presence; uses the wan2.1 VAE
• wav2vec2 audio encoder model class, backend module, --audio-encoder flag
• S2V modules in the Wan DiT: causal audio encoder, motion encoder, audio injector (AdaLN + cross-attention), reference-latent tokens
• audio windowing/bucketing to the latent frame count; uncond uses silence
• new --audio flag; mux the driving track into avi/webm (wav sidecar)
• skip ComfyUI fp8_scaled .scale_input keys at parse time
• docs: downloads, S2V example, sampler and 24 GB VRAM guidance

Additional Information

Verification

wav2vec2 audio encoder parity (vs ComfyUI torch reference):

• Final hidden state max abs error 0.0082 against the fp32 torch reference, inside the torch-fp16-vs-fp32 envelope (torch's own fp16 execution error: 0.0221).
  Per-layer error ratio 0.35-0.94 across layers, max 1.02 at layers 22/23.
  Includes the grouped positional conv (weight-norm, drop-last) and stable-LN conv frontend.
  A deliberately broken negative control fails the gate (0.547 vs 0.02 tolerance), so the harness is not vacuous.

End-to-end (RX 7900 XTX 24 GB, ROCm, GGML_HIP_GRAPHS=OFF, 640x368, 13 frames):

checkpoint	VRAM	s/it
wan2.2_s2v_14B_int8_convrot.safetensors	17.6 GB	5.05
wan2.2_s2v-14B-Q8_0.gguf	18.4 GB	6.19
wan2.2_s2v-14B-Q4_K_M.gguf	13.4 GB	10.99

• Full 77-frame run (int8_convrot + lightx2v low-noise lightning LoRA, 4 steps, cfg 1.0): ~164 s wall including ~50 s VAE decode.
• Baseline quality run (int8_convrot, 20 steps, cfg 6.0, no LoRA, 33 frames): 14.3 s/it.
• Output audio muxing validated end-to-end on ROCm and Vulkan (driving track copied and truncated to video duration, e.g. 236147 -> 13000 samples for a 13-frame clip, muxed into .avi).

Note: int8_convrot only works on CUDA out of the box with the current ggml revision.
ROCm and Vulkan support comes from my ggml PRs (leejet/ggml#6-#9, currently open).
The ROCm/Vulkan numbers above were measured with that stack applied. GGUF checkpoints (Q8_0 / Q4_K_M) are unaffected by any of this.

Regression (existing Wan models unchanged):

• Wan2.1 T2V 14B (Q8_0 GGUF, ROCm)
• Wan2.1 I2V 14B 480p (Q8_0 GGUF + clip_vision, Vulkan)
• Wan2.2 TI2V 5B T2V + I2V (fp16, 48-ch VAE, ROCm)
• Wan2.2 T2V A14B + I2V A14B (int8_convrot MoE, ROCm)

All report the correct Version: lines and never detect as S2V.

Known limitations

• One generation covers the first S2V chunk window (--video-frames, 77-81 frames at 16 fps); long-video chunked extend mode is not implemented yet.
• bf16 and fp8_scaled checkpoints do not fit 24 GB GPUs. int8_convrot, Q8_0, Q4_K_M tested ok.
• On ROCm, tiled VAE decode of longer clips currently needs GGML_HIP_GRAPHS=OFF (HIP graph capture issue for that graph shape, not specific to this PR's code paths).

Sample outputs

77 frames, 512x512, 56.86s/it, ROCm, generate_video completed in 1176.76s

s2v_test_q22_h264.mp4

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

LLM Disclosure: Yes, used GLM-5.3-Flash