Implement deferred imports

[smcolby]

Description

Importing openadmet.models.registries was paying the full cost of every 3rd-party library in the package (~6.7s cold) regardless of which components were actually needed. This PR replaces the "import everything at once" paradigm with deferred/lazy imports across all four component groups (models, featurizers, splitters, trainers/evaluators). Resolves #550 .

Results

Benchmark	Before	After	Speedup
import openadmet.models.registries	6.702s	0.111s	60×
import openadmet.models	0.044s	0.017s	2.6×
registries + load_all()	6.702s	3.727s	1.8× (+ fully deferrable)
architecture/model_base.py	1.652s	0.070s	24×
architecture/chemprop.py	3.083s	0.101s	30×
split/cluster.py	3.524s	0.331s	11×
trainer/lightning.py	1.653s	0.069s	24×
eval/regression.py	1.582s	0.326s	4.9×

Changes

Split model_base.py

• Extracted LightningModuleBase and LightningModelBase into a new architecture/lightning_model_base.py. All torch / lightning imports are isolated there.
• model_base.py uses PEP 562 module __getattr__ to lazily re-export the Lightning classes, preserving backward-compatible from model_base import LightningModelBase without paying the torch cost at import time.
• Deferred joblib inside save() / load().

Deferred estimator class imports

• Replaced mod_class: ClassVar[type] = SomeThirdPartyClass with a _get_estimator_class() classmethod in every concrete pickleable model (xgboost, catboost, lgbm, rf, svm, tabpfn, dummy). Each classmethod contains a local import that fires only at first build() call.
• Moved _METRIC_TO_LOSS dict from module level into chemprop.build().

Deferred imports in features / split / trainer / eval

• features/feature_base.py: heavy molfeat / torch imports moved to TYPE_CHECKING block; from __future__ import annotations added.
• features/chemprop.py: removed a self-import bug (lines importing from itself); all chemprop / torch / sklearn imports deferred inside featurize() and _vendor_build_dataloader().
• split/scaffold.py: splito and sklearn.model_selection.train_test_split deferred inside each split() method.
• split/cluster.py: useful_rdkit_utils, datamol, molfeat, KMeans deferred inside split(); removed unused GroupShuffleSplit import.
• trainer/lightning.py: torch, lightning, and all callbacks/loggers deferred inside build() / train().
• eval/regression.py: wandb deferred inside if self.use_wandb: blocks; scipy.stats, sklearn.metrics, and seaborn deferred by converting the class-level _metrics dict into a _base_metrics() classmethod and moving plot imports inside plot methods.
• eval/eval_base.py: scipy.stats.bootstrap deferred inside stat_and_bootstrap().
• eval/cross_validation.py: stopped importing removed module-level names from regression.py; wrap_ktau / wrap_spearmanr now do local imports.

Lazy registry loading

• New openadmet/models/_registry_loader.py exposes load_group(name) and load_all() (both idempotent), using importlib.import_module. Zero heavy imports at module level.
• registries.py rewritten to only import the six registry objects (models, featurizers, splitters, trainers, evaluators, ensemblers) plus re-export load_all.
• Every get_*_class() function now calls load_group() before the registry lookup, so any single-component usage auto-loads only what it needs.
• anvil/specification.py and anvil/workflow_base.py updated to import load_all instead of from registries import *.

Quality Assurance & AI Policy

To maintain project quality and respect maintainer bandwidth, please confirm the following:

• Manual Verification: I have manually reviewed and tested the code in this PR.
• AI-Assisted Content: If AI tools were used (e.g., Copilot, ChatGPT), I have personally verified the logic, edge cases, and compliance with the existing codebase. I confirm the code is not a "blind" AI generation.
• Minimal Review: I believe this PR is in a state that requires minimal intervention or correction from maintainers.
• Scoped Change: This PR addresses a single, well-scoped issue rather than multiple unrelated changes.

Status

• Ready to go (Checking this signals to maintainers that the PR is ready for final review)

Developers Certificate of Origin

• I certify that this contribution is covered by the MIT License here and the Developer Certificate of Origin at https://developercertificate.org/.

---

Note to Contributors: We reserve the right to close PRs without review if they appear to lack human validation or do not meet the quality standards described in our CONTRIBUTING.md.