arrow

ARROW-4490: [Rust] Add explicit SIMD vectorization for boolean ops in "array_ops"

This PR adds explicit SIMD for boolean ops, `and`, `or` and `not`.

I moved `array_ops` into the new `compute` module.  From the outside this module serves the same purpose as the previous `array_ops` module (all kernels will be accessible from this namespace) and the remaining `array_ops` implementations are exposed via the `compute` module currently.  As I add explicit SIMD for more kernels they will migrate from `array_ops` into their own modules under `compute`.  I am keeping sub-modules under `compute` (as apposed to compute.rs) as SIMD can get rather verbose and it seems that `compute` may expand in the future.

I have included benchmarks where I re-create the old default implementations for others to take a look at the speed improvement.  It's not clear whether we need the non-SIMD versions in the benchmarks long term but I left them in for now to make the non-SIMD/SIMD comparison.

There are likely more optimizations possible (processing the values and null bit buffers in a single loop for instance) but I wanted to get the cleanest impl first and add further optimizations later if needed.

Author: Paddy Horan <paddyhoran@hotmail.com>

Closes #3641 from paddyhoran/boolean-kernels and squashes the following commits:

89588e6 <Paddy Horan> Removed `compute` from `mod.rs`
f9ae58a <Paddy Horan> Updated benchmarks
e321cec <Paddy Horan> Updated `not` to use trait impls
da16486 <Paddy Horan> Implemented `Not` for `Buffer`
f21253d <Paddy Horan> Updated datafusion and comments
a3c01c8 <Paddy Horan> Added SIMD binary boolean kernels