Shift resized intervals to stay on chromosome

[buyan-kh]

Summary

grelu.sequence.utils.resize can produce start < 0 or end > chrom_size when applied to intervals (for example adata.var = resize(adata.var, seq_len)). That failure only shows up later, when a dataset calls check_chrom_ends.

This adds an optional genome= argument on the intervals path. When it is set, chrom sizes are loaded with grelu.io.genome.read_sizes and windows that leave the chromosome are shifted so they stay on-chromosome while keeping seq_len. Rows are never dropped, so AnnData X/var alignment is preserved. If seq_len is larger than a chromosome, resize raises. genome=None (the default) keeps the previous unconstrained behavior.

This is intended for #176. It does not change the open column-order work on #198.

Compare for an upstream PR: https://github.com/Genentech/gReLU/compare/main...buyan-kh:gReLU:cursor/resize-genome-bounds-300d?expand=1

Fixes #176

Test plan

Ran locally (this environment does not have the full gReLU stack / hg38):

PYTHONPATH=src python3 -m pytest tests/test_sequence.py::test_resize tests/test_sequence.py::test_resize_intervals_genome -o addopts=

2 passed in 1.39s

Also ran the rest of tests/test_sequence.py excluding hg38 / tangermeme tests:

PYTHONPATH=src python3 -m pytest tests/test_sequence.py -o addopts= -k "not test_seq_formatting and not test_gc_distribution and not test_dinuc_shuffle"

13 passed, 3 deselected in 1.24s

Coverage of the new tests (tests/files/test_genome.fa / .sizes, chr10/chr21 size 2000):

• OOB-left: start < 0 is shifted to [0, seq_len)
• OOB-right: end > chrom_size is shifted to [size - seq_len, size)
• seq_len larger than the chromosome raises
• genome=None still allows out-of-bounds coordinates
• row count / index preserved (no drops)