The package worked for the happy path but had several defects that only
appeared at the edges, no CI on commits, and a README that stopped at a
single example. This reworks all three while keeping the 2.0.x API
working.

Correctness fixes:

- Target(target_value=0) was discarded. The default was chosen with
  `target_value or <bound>`, so a falsy 0 became an unreachable bound and
  the session could never stop on that target.
- last_value raised AttributeError before the first round, and after an
  interrupt during it. It is now derived from the persisted history, so it
  is also correct straight after a restart.
- Target() was a mutable default argument shared by every trainer.
- train()/predict before compile() raised AttributeError on an internal
  attribute; they now raise RuntimeError explaining what to call.
- A target naming an absent metric raised a bare KeyError; it now lists the
  available keys and points at validation_data for val_* names.
- Checkpoint parent directories were not created, so nested paths failed.
- Restored values stayed as 0-d object arrays instead of floats/lists.
- requires-python said >=3.9, but TensorFlow has needed >=3.10 since 2.16.
- Dropped the unused polars dependency.

Structure and tooling:

- Split into target.py, trainer.py and _storage.py; public imports unchanged.
- Renamed InfinityTraining -> InfiniteTrainer and optimize_* -> best_*, with
  the old names kept as aliases that emit DeprecationWarning until 3.0.0.
- Added a CI workflow running lint, tests on 3.10-3.12 and a packaging check
  on every push and PR. Tests previously ran only when a release was cut.
- The publish workflow now verifies the tag matches the packaged version.
- Moved run.py to examples/mnist.py so it no longer ships in the wheel.

Tests and docs:

- 55 tests, up from 3, covering targets, stop conditions, persistence and
  resume, error handling, and the deprecation shims.
- Rewrote the README with an API reference, a resume guide, a timeout
  caveat and a checkpoint security note; added CHANGELOG.md and
  CONTRIBUTING.md.

Verified: 55 passed, ruff check and format clean, twine check passed, and
the README quick-start and resume flows run as written.

Raising target-version to py310 enabled B905, and zip() over weight lists
is exactly where an unnoticed length mismatch would matter, so pass
strict=True rather than silencing it. Also import Callable from
collections.abc.