New models and refactor around proc_data.py

[Cloud0310]

Summary

This PR refactors the Python data pipeline and adds configurable MobileNet model training.

Data Pipeline

• Split the previous proc_data.py implementation into packaged modules under python/detypify.
• Store raw LaTeX-labelled datasets separately, then map labels to Typst symbols locally with datasets caching.
• Generate frontend metadata under build/generated.
• Update the weekly Typst metadata workflow to compare effective mapping digests before regenerating metadata.

Training And Models

• Replace the fixed model enum with mobilenet_{v4|v5}_{size} model names.
• Add MobileNetV5 support as experimental.
• Add --compile/--no-compile for torch.compile.
• Add CPU/CUDA/RoCM dependency extras for training environments.
• Remove TensorRT dependencies and options.

Frontend And CI

• Update frontend service metadata imports to use generated metadata.
• Add frontend dependency security updates.
• Move Ruff configuration into ruff.toml.

Notes

• MobileNetV5 support is experimental and needs training validation.
• Training should use explicit accelerator extras, for example uv run --extra cuda python/train.py.

[Cloud0310]

@QuarticCat Please try new model training, offer me the test set result image and training log. All should be under dir build.

Here's the some things needs your notification when training:
When training, add --no-ema cli option, as I want just the raw exp data.
Models to be trained: v4_035, v5_010, v_005. (the number means percentage of full size model. here, is 35% of small_v4 variant, 10% of full size v5).