Formulaic

This is a minor patch release that fixes an attempt to import `numpy.typing` when numpy is not version 1.20 or later. (thanks for noticing this and fixing it bashtage ).

This is a minor patch release that fixes the maintaining of output types, NA-handling, and assurance of full-rank for factors that evaluate to pre-encoded columns when constructing a model matrix from a pre-defined `ModelSpec`. The benchmarks were also updated.

This is a relatively massive release bringing lots of new features and bugfixes. All users are encouraged to upgrade. The full changelog can be viewed in the [documentation website](https://matthewwardrop.github.io/formulaic/changelog/). The highlights include:

**Breaking changes:**

* The minimum supported version of Python is now 3.7 (up from 3.6).
* Moved transform implementations from `formulaic.materializers.transforms` to
the top-level `formulaic.transforms` module, and ported all existing
transforms to output `FactorValues` types rather than dictionaries.
`FactorValues` is an object proxy that allows output types like
`pandas.DataFrame`s to be used as they normally would, with some additional
metadata for formulaic accessible via the `__formulaic_metadata__`
attribute. This makes non-formula direct usage of these transforms much more
pleasant.
* `~` is no longer a generic formula separator, and can only be used once in a
formula. Please use the newly added `|` operator to separate a formula into
multiple parts.

**New features and enhancements:**

* Added support for "structured" formulas, and updated the `~` operator to use
them. Structured formulas can have named substructures, for example: `lhs`
and `rhs` for the `~` operator. The representation of formulas has been
updated to show this structure.
* Added support for context-sensitivity during the resolution of operators,
allowing more flexible operators to be implemented (this is exploited by the
`|` operator which splits formulas into multiple parts).
* The `formulaic.model_matrix` syntactic sugar function now accepts `ModelSpec`
and `ModelMatrix` instances as the "formula" spec, making generation of
matrices with the same form as previously generated matrices more
convenient.
* Added the `poly` transform (compatible with R and patsy).
* `numpy` is now always available in formulas via `np`, allowing formulas like
`np.sum(x)`. For convenience, `log`, `log10`, `log2`, `exp`, `exp10` and
`exp2` are now exposed as transforms independent of user context.
* Pickleability is now guaranteed and tested via unit tests. Failure to pickle
any formulaic metadata object (such as formulas, model specs, etc) is
considered a bug.
* The capturing of user context for use in formula materialization has been
split out into a utility method `formulaic.utils.context.capture_context()`.
This can be used by libraries that wrap Formulaic to capture the variables
and/or transforms available in a users' environment where appropriate.

This is a minor release that fixes an issue whereby the `ModelSpec` instances attached to `ModelMatrix` objects would keep reference to the original data, greatly inflating the size of the `ModelSpec`. (Many thanks to mmacpherson for reporting!)

This release is identical to v0.2.2, except that the source distribution now includes the docs, license, and tox configuration.

This is a minor release with one bugfix.

- Fix pandas model matrix outputs when constants are generated as part of model matrix construction and the incoming dataframe has a custom rather than range index. (issue noted by CamDavidsonPilon)