refactor: fix correctness bugs, add tests, and document the package

[vyncint]

Review, refactor and documentation pass over the package. Backward compatible with 2.0.0 — existing code keeps working and warns where a name has changed. Ships as 2.1.0.

Why

The happy path worked, but a few edge cases were broken, tests only ran when a release was published, and the README stopped after one example. This addresses all three.

Correctness fixes

Problem	Effect	Fix
target_value or <bound>	Target(target_value=0) silently became an unreachable bound, so training could never stop on that target	Test for None
last_value only assigned inside the training loop	AttributeError before the first round, after an interrupt during it, and on every resumed session	Derived from the persisted history
target: Target = Target()	One mutable default shared by every trainer instance	None sentinel
No guard before compile()	AttributeError on an internal attribute	RuntimeError naming the missing call
Absent metric name	Bare KeyError	RuntimeError listing available keys, hinting at validation_data for val_*
Parent dirs not created	Nested checkpoint paths failed	Created on save
Values restored as 0-d object arrays	Odd types leaking into user code	Converted to float / plain weight lists
requires-python = ">=3.9"	TensorFlow has required >=3.10 since 2.16, so the floor advertised an uninstallable combination	Corrected to >=3.10
polars dependency	Never imported	Removed

Structure

• Split into target.py, trainer.py, _storage.py. Public imports unchanged.
• InfinityTraining → InfiniteTrainer, optimize_* → best_*. Old names, attributes, constructor keywords and predict_optimize() all still work and emit DeprecationWarning until 3.0.0 — pinned by tests/test_deprecations.py.
• run.py → examples/mnist.py, so it no longer ships inside the wheel.
• Stop reasons reported through logging; Ctrl+C is no longer swallowed silently.

CI

Tests previously ran only on release publication, so nothing gated a commit. Added ci.yml: ruff lint + format, pytest on Python 3.10/3.11/3.12, and a build + twine check, on every push and PR. The publish workflow now also verifies the release tag matches the packaged version.

Tests

55, up from 3 — targets and comparison rules, stop conditions, best-weight tracking, persistence and resume across sessions, error handling, and the deprecation shims.

Docs

README rewritten with a quick start, a "how it works" walkthrough, a resume guide, a full API reference, a timeout-granularity caveat and a security note about pickled checkpoints. Added CHANGELOG.md and CONTRIBUTING.md.

Verification

pytest tests            55 passed
ruff check .            All checks passed
ruff format --check .   13 files already formatted
python -m build         built sdist + wheel
twine check dist/*      PASSED

The wheel contains only infinite_training/ — no tests, no examples. The README quick start and resume snippets were run verbatim: the quick start reaches its 1e-4 target after 25 rounds, and a fresh trainer on the same paths reports rounds_completed == 25 with last_value and best_value restored.

Notes for review

• The two behaviour fixes users could notice are the target_value=0 change and last_value returning None instead of raising; both are in the changelog.
• requires-python now excludes 3.9. Worth a sanity check that you are happy dropping a version that could not have had a working TensorFlow anyway.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.