Awkward

reikdas has added a second batch of CUDA translations.

<img src="https://user-images.githubusercontent.com/1852447/92425142-f7c5d480-f14b-11ea-8268-87672d6b824b.png" width="600">

nsmith- fixed reference-cycle issues in VirtualArrays by making ArrayCache keep only a weak reference to the cache (that it might be included in).

jpivarski fixed some odds and ends and also revamped the deployment procedure so that `pip install awkward1` doesn't install anything outside of its Python directory. (Homebrew complains about that on MacOS.) There is now a `python -m awkward1.config --cflags --libs` command to get a path to the installed header files and libraries, and `dependent-project` is now being tested regularly on Linux and MacOS.

Quick fix for inability to print out partitioned arrays.

trickarcher added conversions to and from CuPy arrays, enabling all of CuPy's ufuncs for non-structure-changing math (i.e. formulas). (372)

reikdas improved the kernel testing (412, 413, 414).

jpivarski fixed references to `np.object` (where `np` is `NumpyMetadata`, not `numpy`), which was breaking Uproot4 tests and motivated this new release.

jpivarski also fixed 415, issues with `ArrayBuilder.append` and union types.

veprbl fixed the warnings in Linux (gcc) and MacOS (clang): everything is clean!

jpivarski added a categorical type with functions like `ak.to_categorical`. This type is high-level only, like the string type, and is used to determine which fields to convert to Arrow/Parquet's dictionary encoding and which not to. See issue 400.

Fixed 408, a severe error in `ak.argmin` and `ak.argmax`: they returned wrong integer positions for arrays with missing values or variable-length lists with `axis != -1`.

nsmith- filled in missing data in VirtualArrays so they can be used in Numba in more situations.

reikdas cleaned up the kernel tests.

jpivarski fixed some bugs in the Numba implementation, enabled functions like `enumerate` and `zip`, added GitHub URLs to all exceptions for faster investigation of bugs, and abstracted the use of NumPy so that the Python-side calculations can be swapped for CuPy.

Updated the awkward1-cuda-kernels build and testing process. This is the first regular release of awkward1-cuda-kernels as a PyPI package, and henceforth all awkward1 PyPI releases will be accompanied by a awkward1-cuda-kernels PyPI release with the same version number.