Training: DataLoader, feature cache, AMP, lazy package exports

[cropsgg]

Summary

Updates training/inference plumbing so pipelines consume AppSettings: batched DNABERT encoding, optional on-disk feature cache, DataLoader-based training, optional AMP, and lazy imports in bloom_dnabert/__init__.py so importing light modules does not require every heavy dependency at import time.

Base branch: pr/02-clinvar-data-loader (stacked — merge after #2 or retarget once earlier PRs land).

What changed

• bloom_dnabert/feature_cache.py — Fingerprinted cache for precomputed Bloom/DNABERT features when data.feature_cache_dir is set.
• bloom_dnabert/classifier.py — HybridClassifierPipeline and BloomGuidedPipeline read batch sizes, workers, AMP, max length, etc. from settings.
• bloom_dnabert/dnabert_wrapper.py — Batch encode / token-level outputs aligned with training loops.
• bloom_dnabert/bloom_filter.py — Aligns with config paths and seed loading APIs used by the app/CLI.
• bloom_dnabert/__init__.py — Lazy __getattr__ exports for torch/transformer-heavy submodules.

Why this PR exists

Scales experimentation beyond notebook-style loops: reproducible batching, less redundant DNABERT work via cache, and cleaner import graph for tests and partial installs.

Dependencies

• PR Data: config-driven ClinVarDataLoader and reference-validated windows #2 (Data: config-driven ClinVarDataLoader and reference-validated windows #2): data loader + settings-aware data path.

How to verify

Run classifier/bloom tests in an environment with torch and dependencies from requirements.txt (full pytest tests/).