hotfix to remove local path from uv source in pyproject.

updates infer script to fix a bug, where models trained on X_hvg would incorrectly use a randomly initialized "gene deoder MLP". this would effectively randomize the predictions of the model

note that this only occurred with src/state/_cli/_tx/_infer.py and did not affect src/state/_cli/_tx/_train.py or src/state/_cli/_tx/_predict.py scripts

Updates code to allow:

data.kwargs.output_space == "embedding" to train on only embeddings without a gene decoder
update tx predict to clip [0,14]
update model hyperparameters for better preset values

several improvements in this overdue change log.

1. infer now correctly groups control cells by shared covariates. see here for examples of running inference: https://colab.research.google.com/drive/1bq5v7hixnM-tZHwNdgPiuuDo6kuiwLKJ?authuser=1#scrollTo=ENfnF6ofAz1M

2. the embedding model checkpoints are now completely packaged (so no more model folder is needed). we retain backwards compatibility. here is an example for running transform: https://colab.research.google.com/drive/1uJinTJLSesJeot0mP254fQpSxGuDEsZt

3. the auxillary files to train the SE model, as well as the complete training data, are available here: https://huggingface.co/datasets/arcinstitute/SE-167M-Human

some notes on how to train SE yourself: #138 (comment)

4. new eval scripts for SE checkpoints: use uv run state emb eval

this version fixes a bug in the repo .gitignore that doesn't distribute the wandb folder for the configs.

the model is trained in bf16-mixed precision, so this patch uses bfloat16 for inference when cuda is available. empirically for the 600m model this improves inference speed ~8-9x

fix for fresh install to work with or without wandb integration

adds example command

this upgrades the version of cell load so that the perturbed cell and control cell barcodes are available in the final predicted outputs