Compatibility fixes for modern NumPy / sklearn / HMMER / CheckM

[rec3141]

Summary

Update COMEBin to run with modern dependencies (Python 3.11, NumPy 1.26, scikit-learn 1.4+, PyTorch 2.2+), replacing removed/deprecated APIs and fixing env solver failures.

What changed

Python code fixes:

• sklearn.cluster._kmeans private imports → public APIs (KMeans, euclidean_distances, check_random_state, row_norms)
• KMeans(algorithm="full", n_jobs=-1) → KMeans(algorithm="lloyd", n_init=10)
• stable_cumsum → np.cumsum(dtype=np.float64)
• np.int → np.int64 (5 instances across 3 files)
• .item(0) → .item() (deprecated positional arg in NumPy 2.x)
• torch.cuda.amp.{GradScaler,autocast} → torch.amp.{GradScaler,autocast} with explicit device_type (PyTorch 2.4+)
• ruamel.yaml Path loading → file handle (avoids deprecation warning)
• hmmsearch: fallback to -E 1e-10 when --cut_tc fails (HMMER ≥3.3)

Runtime hardening:

• Auto-resolve CHECKM_DATA_PATH from $CONDA_PREFIX
• Auto-set MPLCONFIGDIR to output dir (avoids writes to restricted $HOME)
• Lazy imports for lightweight subcommands
• Suppress pkg_resources deprecation warning from CheckM

Environment (comebin_env.yaml):

• Python 3.10–3.11 (pinned <3.12 for CheckM1 compatibility)
• NumPy >=1.26,<2.0, scikit-learn >=1.4, PyTorch >=2.2
• Moved hnswlib, python-igraph, leidenalg, scanpy from pip to conda-forge
• Pinned pplacer=1.1.alpha19 (>=1.1 doesn't match alpha version strings → solver failure)
• Added missing ruamel.yaml and tqdm dependencies
• CPU-only PyTorch via conda-forge (GPU users can override)

Issue tags

Fixes #44
Refs #29, #38, #11

Validation

• Environment solves cleanly with mamba env create -f comebin_env.yaml
• All COMEBin Python modules import without errors
• sklearn, torch.amp, numpy API smoke tests pass
• End-to-end run with real contigs + BAMs (pending)

AI Assistance

Portions of this PR were developed with assistance from OpenAI Codex (initial refactoring) and Claude Code (code review, additional fixes, env debugging).