Repackaging RISE

[aarmey]

No description provided.

[aarmey]

@copilot summarize this PR.

[Copilot]

@copilot summarize this PR.

This PR repackages RISE as a proper Python package and includes two sets of changes:

Overall structure (initial commit):

• Establishes RISE as an installable Python package (pyproject.toml, hatchling build, MIT license) with a clean RISE/ module containing factorization and plotting submodules
• Adds CI/CD workflows for building, testing, and publishing documentation
• Includes a full Sphinx documentation site with tutorial, API reference, and tutorial images
• Ships analysis scripts, figure generation code, and Jupyter notebooks for the Lupus and Thomson PBMC datasets
• Adds pyproject.toml-managed dependency groups (analysis, docs, dev) with uv.lock for reproducibility

Memory usage fixes (second commit):

• Lupus loading: Replaces anndata.read_h5ad(..., backed="r") with direct h5py reads of only the needed HDF5 elements (obs, raw/var, raw/X). The old approach caused ~180 GB of decompression buffer allocation even in backed mode because hdf5plugin pre-allocated buffers for the dense processed X slot (9.6 GB float32).
• OOM on subsetting: Adds .copy() after the boolean-mask filter before passing to prepare_dataset. Without it, prepare_dataset's X.X = csr_array(X.X) triggered a setitem on the parent sparse matrix (1.26M × 32,738 CSR → LIL/dense conversion → OOM).
• Makes the Thomson gating import lazy (inside the function body) to avoid loading unnecessary modules at import time.