Awkward

Renamed and fixed ak.values_astype, added Pandas-style ak.Array constructor, and other touch-ups as part of writing documentation.

nsmith made ak.from_arrayset "even more lazy" by inserting a VirtualArray in every field of a RecordArray when `lazy=True`.

The website can't be updated unless the bug-fixes that the tutorials require are available as a released version. Thus, this is just a bug-fix release (there may be more, as more documentation gets written).

As per 350, Pandas integration will be deprecated. Starting in version 0.3.0, it won't be possible to use Awkward arrays as Pandas Series or DataFrame columns. We're also giving up on attempts to keep Matplotlib from iterating over non-numeric data. (These are part of a theme of removing sneaky tricks from the code, such as `awkward1._util.called_by_module`, for robustness.)

Updated to backward-incompatible changes in Arrow 1.0.0. Now Arrow 1.0.0 is the minimum version that works with Awkward.

Updated to changes in JupyterBooks to fix documentation.

ianna added a `numbers_to_type` function, which converts all numbers to a given type while maintaining an array's structure.

reikdas generated over three thousand unit tests for the kernels, which previously had only been tested as part of integration tests. Once the blacklist in the test generation framework is reduced to zero, line coverage of kernels will be 100%. Once the samples are carefully chosen in kernel-specification/samples.json, all the important cases will be tested as well. This testing framework will be extended to CUDA kernels as well.

nsmith- added class and class method decorators to simplify the process of adding high-level behaviors.

Internally, all of the "offset" parameters have been removed from kernel argument lists. Instead, correctly offset pointers are passed to these functions, so they can assume that they start from (relative) array index zero. Several previously fixed bugs have been traced to this and in the conversion, I found a few more suspicious spots—now we don't have to worry about it anymore.

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

Make Awkward-in-Pandas an opt-in feature with likely deprecation, as described in 350. (That's where users can explain their use-cases if they want to keep it.)

Change `metadata` with a `"cache"` into just `cache`, since "metadata" is a legitimate word for scientific data and we don't want a conflict. (It broke Coffea's daily tests.)

This version includes trickarcher's first full example of moving an array to the GPU and computing something on it (ak.num). The necessary code does not exist in PyPI yet, though.

From nsmith-, docstrings are now passed through array manipulations as a `__doc__` parameter (which becomes a `__doc__` attribute on `ak.Array`).

ianna implemented negative `axis` parameters, which includes cases like::

array = ak.Array([
{"x": [1], "y": [[], [1]]},
{"x": [1, 2], "y": [[], [1], [1, 2]]},
{"x": [1, 2, 3], "y": [[], [1], [1, 2], [1, 2, 3]]}])

ak.num(array, axis=-1).tolist() == [{'x': 1, 'y': [0, 1]},
{'x': 2, 'y': [0, 1, 2]},
{'x': 3, 'y': [0, 1, 2, 3]}]

which would be inexpressible with a nonnegative `axis`.

Sorting finally has a high-level function, `ak.sort` (and `ak.argsort`).

Awkward ↔ Apache Arrow conversion now prefers 32-bit Arrow types, if possible, and Awkward arrays can be written to and read from Parquet files.

Awkward Arrays can now be deconstructed into a Mapping of str → NumPy arrays (or binary blobs) and back again using `ak.to_arrayset` and `ak.from_arrayset`. This permits general serialization, though standard formats like Arrow and Parquet are to be preferred for any long-term storage.

Added sort/argsort; kernels are being refactored and documented to prepare for GPU support; performance tuning: RecordArray carry → IndexedArray; `__getitem__` for masked, jagged slices is now correct.