- Classes instead of functions for the `IntParam`, `StrParam`, etc.
- Strong signature for the `ConfigParam` type
- `filter` was renamed to `apply`, because that's closer to its functionality
- `is_valid` was renamed to `validate` to match `apply`
- Both `apply` and `validate` callables can now be set for all params
- `DeviceParam` was moved over to where the other config param types are defined
- Already deprecated config parameters were removed
- Add a few more tests

This way the "convert" kwarg can be removed from the signature of the EnumStr
param

Co-authored-by: Marco Gorelli <m.gorelli@samsung.com>

Reconfigure codecov to reduce false alarm rate.

Remove Variable.fgraph and Variable.clients properties

This change requires the addition of an `fgraph` argument to the
`Op.infer_shape` method used by `ShapeFeature`.

These tests made unnecessary assumptions about results derived from unrelated
optimizations, so any changes to the those optimizations would cause the tests
to fail--even when the actual functionality being tested would succeed.

This commit removes uses of `pre_greedy_local_optimizer` and
`pre_constant_merge` that needlessly introduce complexity into the
optimization process through in-place graph changes that occur completely
outside of the optimization/`FunctionGraph` framework.

Closes #190.

These tests actually confirm that `pre_greedy_local_optimizer` traverses a graph
and applies optimizations, and that it avoids terms in a given `FunctionGraph`.

Refactor tests.tensor.nnet optimization tests

Implement a JAX conversion for the Scan Op