Eliminate class proliferation by making each learner directly accept three
independent ABCs. MLPLearner and LinearLearner now take optional optimizer,
bounder, and normalizer parameters instead of requiring wrapper classes.

- Add Bounder ABC with ObGDBounding (decoupled from optimizer)
- Add AutostepParamState + init_for_shape/update_from_gradient for MLP
- Merge 8 learning loops into 4 (absorbed normalized variants)
- Remove NormalizedLinearLearner, NormalizedMLPLearner, and 6 state types
- Update all tests, examples, and utilities for new composable API
- Update README with new features and API stability disclaimer
- Bump to v0.7.0