refactor: split ak._do into meta and content

[agoose77]

This PR:

1. Splits Content._mergeable_next and ak._do.mergeable into Meta._mergeable_next and ak._do.meta.mergeable.
2. Adds dtype to NumpyMeta
3. Drops use of Generic in type hints for Meta classes

(3) is required because of the lack of higher-kinded types. This means that our type hints will only let us prove that NumpyArray is a NumpyMeta in some contexts, e.g. via ak._meta.meta.is_numpy()

[codecov]

Codecov Report

Attention: Patch coverage is 92.19858% with 44 lines in your changes are missing coverage. Please review.

Project coverage is 81.95%. Comparing base (123fa09) to head (7d62858).
Report is 107 commits behind head on main.

❗ Current head 7d62858 differs from pull request most recent head dc9645f. Consider uploading reports for the commit dc9645f to get more accurate results

Additional details and impacted files

Files	Coverage Δ
src/awkward/__init__.py	97.14% <100.00%> (+0.08%)	⬆️
src/awkward/_connect/jax/trees.py	92.30% <100.00%> (+0.09%)	⬆️
src/awkward/_connect/numpy.py	92.05% <100.00%> (+0.03%)	⬆️
src/awkward/_meta/emptymeta.py	94.11% <100.00%> (+0.36%)	⬆️
src/awkward/_meta/indexedoptionmeta.py	100.00% <100.00%> (ø)
src/awkward/_meta/unionmeta.py	94.11% <100.00%> (+0.07%)	⬆️
src/awkward/_nplikes/array_module.py	86.28% <ø> (-0.10%)	⬇️
src/awkward/_nplikes/numpy_like.py	75.49% <100.00%> (+0.19%)	⬆️
src/awkward/_nplikes/typetracer.py	74.61% <ø> (-0.06%)	⬇️
src/awkward/_pickle.py	85.71% <100.00%> (-1.59%)	⬇️
... and 141 more

[jpivarski]

I see that this splits the ak._do module into two submodules for Content and the superclasses of both Content classes and Form classes.

Meta replaces the TypeVar T in many places; I suppose that's because T used to represent a Content or a Form, which had no common superclass, and now it can all be OOP-style, rather than template-style.

The merge* functions have moved. That's fine.

On the whole, there's a lot of differences, but it seems to be just (1) moving things and (2) merging the Content and Form hierarchies. Does this PR complete the merging of those hierarchies, or is there another after this?

Also, what will the hierarchy look like after merging is complete? Is it like this?

flowchart LR
    C(Content)
    F(Form)
    M(Meta)
    ListArray --> C
    NumpyArray --> C
    ListForm --> F
    NumpyForm --> F
    ListMeta --> M
    NumpyMeta --> M
    ListArray --> ListMeta
    NumpyArray --> NumpyMeta
    ListForm --> ListMeta
    NumpyForm --> NumpyMeta

[agoose77]

Meta replaces the TypeVar T in many places; I suppose that's because T used to represent a Content or a Form, which had no common superclass, and now it can all be OOP-style, rather than template-style.

The T was used when I was trying to preserve the idea that ListOffsetArray.content is a Content, not just a Meta. But this ends up needing higher kinded types, which mypy does not yet express.

[agoose77]

On the whole, there's a lot of differences, but it seems to be just (1) moving things and (2) merging the Content and Form hierarchies. Does this PR complete the merging of those hierarchies, or is there another after this?

Yes, I think so. We may be able to promote some Content functions to Meta, but on balance I think its complete.