Zero-copy buffer protocol data import

[kylebarron]

Change list

• Add AnyBufferProtocol enum, with variants for all the Element types. In the FromPyObject impl on AnyBufferProtocol, we simply check for each Element type.
• Implement into_arrow_array for zero-copy conversions of buffer protocol objects. This relies on the Buffer::from_custom_allocation method in arrow. This is safe because the original PyBuffer is put into an Arc and tracked by the Arrow buffer. So when the Arrow array is dropped, it will decrement the reference count on the Arc, and when it hits zero the Arc<PyBuffer> will be dropped. This will at that point call the original buffer's release method.
• Check for the buffer protocol by default in the FromPyObject impl on PyArray.
  • This means that all places in the Python API that accept a PyArray now also accept buffer protocol input
• Remove custom vendoring of PyAnyBuffer from Support for python buffer protocol #156. This was a lot of low-level, relatively unsafe code, that is not ideal to maintain here.
• Add ArrayInput type union, which is a union of ArrowArrayExportable and buffer protocol objects.
• Update type hints to use ArrayInput instead of ArrowArrayExportable where appropriate
• Update README to note zero-copy buffer protocol support

We're now able to do computations on numpy arrays by default!

import numpy as np
import arro3.compute as ac

np_arr = np.arange(10_000_000, dtype=np.uint64)

%timeit np.max(np_arr)
# 1.37 ms ± 46 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

%timeit ac.max(np_arr)
# 1.69 ms ± 32.1 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

This really isn't bad! Especially as numpy might be doing SIMD here (the arrow crate is SIMD capable, but we aren't enabling SIMD optimizations in our wheel builds).