ARROW-10914: [Rust] Refactor simd arithmetic kernels to use chunked iteration

[jhorstmann]

This should avoid out of bounds reads in those kernels and also removes some unsafe blocks. The performance actually seems to increase slightly, I will post the numbers later.

The division kernel looks a bit more complex now, but follows the same pattern as used in the simd aggregation functions, with two separate codepaths for the nullable/non-null cases.

The comparison kernels might also do some out of bounds reads, I will look at those in a separate ticket/PR.

[github-actions]

https://issues.apache.org/jira/browse/ARROW-10914

[jhorstmann]

I'm seeing bigger variations on the benchmark numbers between runs, the results below between current master and this pr are from one of the faster runs:

~/arrow/rust/arrow$ cargo +nightly-2020-11-24 bench --features simd --bench arithmetic_kernels
Compiling arrow v3.0.0-SNAPSHOT (/home/ubuntu/arrow/rust/arrow)
Finished bench [optimized] target(s) in 1m 39s
Running /home/ubuntu/arrow/rust/target/release/deps/arithmetic_kernels-4b00176b1c415f0c
Gnuplot not found, using plotters backend
add 512 time: [319.21 ns 319.38 ns 319.52 ns] change: [-24.243% -24.193% -24.140%] (p = 0.00 < 0.05)
Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
2 (2.00%) low severe
5 (5.00%) low mild
2 (2.00%) high mild
1 (1.00%) high severe
subtract 512 time: [387.54 ns 387.86 ns 388.40 ns] change: [-19.909% -19.766% -19.578%] (p = 0.00 < 0.05)
Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
1 (1.00%) low mild
3 (3.00%) high mild
2 (2.00%) high severe
multiply 512 time: [515.01 ns 517.68 ns 522.09 ns] change: [-9.9324% -9.6166% -9.1766%] (p = 0.00 < 0.05)
Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
1 (1.00%) high mild
3 (3.00%) high severe
divide 512 time: [430.74 ns 430.80 ns 430.86 ns] change: [-41.347% -41.234% -41.143%] (p = 0.00 < 0.05)
Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
1 (1.00%) low mild
6 (6.00%) high mild
3 (3.00%) high severe
limit 512, 512 time: [107.84 ns 107.94 ns 108.12 ns] change: [+0.6778% +0.9025% +1.2602%] (p = 0.00 < 0.05)
Change within noise threshold.
Found 7 outliers among 100 measurements (7.00%)
1 (1.00%) low mild
4 (4.00%) high mild
2 (2.00%) high severe
add_nulls_512 time: [383.53 ns 383.61 ns 383.69 ns] change: [-20.450% -20.350% -20.270%] (p = 0.00 < 0.05)
Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
1 (1.00%) low severe
1 (1.00%) low mild
3 (3.00%) high mild
5 (5.00%) high severe
divide_nulls_512 time: [543.48 ns 543.96 ns 544.84 ns] change: [-23.819% -23.700% -23.587%] (p = 0.00 < 0.05)
Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
3 (3.00%) low severe
1 (1.00%) low mild
1 (1.00%) high mild
5 (5.00%) high severe

[jorgecarleitao]

Thanks a lot, @jhorstmann !

I did not go through all the details of this PR, but what I read looks really good and makes the code much easier to read. So, 👍