Awkward

Broadcasting ufuncs across records is now deprecated, and will be removed in version 1.0.0 (December 1, 2020).

Fixed two bugs (501 and 499).

Supports Python 3.9. (Now MacOS only supports Pythons 2.7, 3.7-3.9 (3.6 dropped). All other operating systems support Pythons 2.7, 3.5-3.9)

Based on [pybind 2.6.0](https://github.com/pybind/pybind11/releases/tag/v2.6.0). This is the first jump in pybind11 version in a while.

Put the `cmake` requirement back into pyproject.toml and version-lock `awkward1-cuda-kernels` to `awkward1`.

The main reason for this release is because we hit the 10 GB limit and now I can't deploy a complete set of wheels. (They won't upload if a name had existed, even if manually deleted.) I've [requested a higher limit](https://github.com/pypa/pypi-support/issues/630).

As discussed in 350, Awkward arrays can no longer be used as Pandas columns. Also part of this API change (464),

* `array.tojson` has been removed: use `ak.to_json(array)`.
* `array.columns` has been renamed: use `array.fields`.
* `ak.keys(array)` has been renamed: use `ak.fields(array)`. (`ak.keys` still exists with a "deprecated until 0.4.0" message)
* `array.ndim` (property) is now the dimension of the array, rather than `1` (a Pandas requirement)
* `array.type` has been added; it is equivalent to `ak.type(array)`. If you had fields named `"type"`, they will now have to be accessed as `array["type"]`, rather than `array.type`.

Thanks to reikdas, this is the first release with a [formal specification](https://github.com/scikit-hep/awkward-1.0/blob/master/kernel-specification.yml) for the cpu-kernels/cuda-kernels library! (The bottom layer of the 3-layer architecture, [documented here](https://awkward-array.readthedocs.io/en/latest/_auto/kernels.html).)

Furthermore, ufunc behavioral matching has been generalized (463) to allow:


ak.behavior[np.multiply, numbers.Number, "Vector"] = scalar_multiply
ak.behavior[np.multiply, "Vector", numbers.Number] = scalar_multiply


Two mistakes were fixed in the Windows builds (465 by chrisburr and 467), allowing them to be used without debug-version system libraries (430). The first of these enabled lgray to [deploy it on conda-forge](https://anaconda.org/conda-forge/awkward1)!

Still built on pybind 2.4.3, though we might want to update that, especially if pybind supports the NumPy datetime dtypes.

Getting closer to finalizing reikdas and trickarcher's work this summer, and nsmith- fixed a missing case in `ak.type` (booleans!).

Also, the Windows build procedure was generalized to allow awkward1 to become a package on conda-forge.

ianna fixed the intermittent MacOS error (issue 320)! Although this bug only manifested itself on MacOS in our tests, due to the capriciousness of memory overwriting/reading, it was really a bug on all platforms.

reikdas auto-generated more CUDA kernels and will now be turning to consolidating what we have (i.e. cleanup). 59 of the 77 embarrassingly parallel kernels can be converted without manual intervention.

jpivarski fixed a bug in ak.flatten and ensured that variable dimensions remain variable and regular dimension remain regular in reducers, unless there are missing values involved. I also added two performance improvements: broadcasting only compacts lists if they have to, resulting in a 40× speedup in the limit of large arrays (PR 444) and when Uproot has types that would be slow to deserialize from ROOT, it can now offload the processing to a C++ loop in Awkward, resulting in an 80× speedup in the targeted case (PR 448). This required ak.concatenate to merge in batches, rather than two at a time (TBaskets in a ROOT file present hundreds of arrays to concatenate; an _O(n²)_ algorithm is bad), which is PR 449 (merged into PR 448 before the latter was squash-and-merged).