[codex] Add Mini Studio web UI and fix LTX config loading

[andyelka-creator]

Summary

• add a lightweight local Gradio Mini Studio for LTX-2 and Wan job launching
• persist queue state, per-job logs, and generated video gallery under outputs/
• add a Web UI guide for local usage and troubleshooting
• fix LTX-2 config loading when Hugging Face metadata keys such as _class_name are present
• normalize unprefixed Hugging Face / Diffusers transformer weight keys before strict MLX loading
• support unified Hugging Face vae/diffusion_pytorch_model.safetensors layout when generation asks for vae/encoder or vae/decoder
• normalize unified Diffusers VAE encoder/decoder key names and map latent_channels into the encoder config
• default missing unified VAE decoder per-channel statistics to identity values for strict loading
• make top-level mlx_video imports lazy so the Web UI can import package metadata without eagerly loading model modules

Fixes #30
Fixes #32
Fixes #33
Fixes #34
Fixes #35
Fixes #36

Root Cause

LTX-2 loading had several compatibility failures against current Hugging Face files:

1. config.json contains metadata keys that are not dataclass fields, so direct LTXModelConfig(**config_dict) fails.
2. The transformer safetensors are unprefixed, but still use Diffusers module names such as linear_1, to_out.0, norm_q, time_embed, and proj_in.
3. The video VAE is stored as unified vae/config.json and vae/diffusion_pytorch_model.safetensors, while generation requested split vae/encoder and vae/decoder directories.
4. Unified VAE safetensors use unprefixed encoder.* and decoder.* Diffusers block names such as resnets/downsamplers/upsamplers, while the MLX modules expect res_blocks and interleaved sampling block indices.
5. Diffusers AutoencoderKLLTX2Video uses out_channels=3 for RGB output and latent_channels=128 for latent width; the encoder must use latent_channels.
6. The unified VAE checkpoint omits optional decoder per-channel statistics, but strict loading still expected those parameters even though the decoder initializes identity defaults.

Validation

• .venv/bin/python -m pytest tests/test_ltx_model_loading.py tests/test_ltx_video_vae_loading.py tests/test_webui_job_store.py tests/test_webui_runner.py tests/test_rope.py::TestDoublePrecisionRopeConfig -q -> 24 passed, 2 warnings
• real downloaded Lightricks/LTX-2 transformer key comparison after sanitizer -> 3510 sanitized keys, 3510 model keys, extra 0, missing 0
• LTXModel.from_pretrained(.../Lightricks/LTX-2/.../transformer, strict=True) -> strict transformer load ok
• VideoEncoder.from_pretrained(.../Lightricks/LTX-2/.../vae/encoder) -> image encode ok, latent shape (1, 128, 1, 16, 16)
• VideoDecoder.from_pretrained(.../Lightricks/LTX-2/.../vae/decoder, strict=True) -> strict decoder load ok, identity statistics shape (128,)
• unified VAE encoder sanitizer against real checkpoint -> extra 0, missing 2 default per-channel statistics
• .venv/bin/mlx_video.webui --help
• curl -sI http://127.0.0.1:7861/ -> HTTP/1.1 200 OK

Local deployment

• The Gradio Web UI is running locally at http://127.0.0.1:7861/ in this workspace.
• This repository does not include a cloud deployment target, so the deployment completed here is the local Web UI service.