docs: record measured VRAM/timings and the two compat deviations

perf: size ZeroGPU duration from measured timings (was 265s for a 52s run)

fix: load weights on CPU then move modules to cuda (ZeroGPU can't intercept safetensors device=cuda)

fix: install write_video shim before vendor import

compat: PyAV implementation of torchvision.io.write_video (removed in tv 0.26)

fix: import ltx_core.loader before ltx_core.quantization (circular import)

Port the authors' multishot_inference_dmd pipeline to ZeroGPU xlarge (fp8-cast DiT, module-scope load)

Upload folder using huggingface_hub

fix: unpin transformers (hf_hub 1.x conflict with gradio 6)

JoyAI-Echo multi-shot AV demo (ZeroGPU, fp8 cast)

initial commit