feat: BMI for bit filtering

[devanbenz]

DRAFT

This commit adds the ability for bit filtering to be done using the _pext_u64 BMI.

Right now this PR is directly copying the bit_util.rs code from: https://github.com/apache/arrow-rs/pull/9848/changes#diff-f2fe5b601534125d0ba033762b62db93ae2f81c71a543b6694df33773ebfa308R939-R976 -- I've just moved the compress function in to arrow-buffer. This draft PR is mostly meant to allow for me to benchmark on my ARM machine and x86_64 machine.

I noticed while bench-marking using the BMI path (filter_bits_compress) leads to a regression when we have a very dense or a very sparse filter mask. Besides those two outliers we gain a large performance gain from BMI.

Case (kept)	strategy	no-BMI2	BMI2	Delta
1023/1024 (dense)	lazy	1.88 us	2.16 us	+15%
	optimized	1.21 us	2.23 us	+84%
	sliced	1.84 us	2.41 us	+31%
9/10	lazy	~103 us	2.98 us	-97%
	optimized	~85 us	3.02 us	-96%
	sliced	~104 us	2.92 us	-97%
1/2	lazy	~39 us	1.94 us	-95%
	optimized	~39 us	1.97 us	-95%
	sliced	~40 us	2.16 us	-95%
1/10	lazy	7.37 us	1.55 us	-79%
	optimized	5.26 us	1.53 us	-71%
	sliced	7.41 us	1.74 us	-77%
1/1024 (sparse)	lazy	0.89 us	0.75 us	-16%
	optimized	99 ns	103 ns	+3.5%
	sliced	0.89 us	0.74 us	-16%

[devanbenz]

This is currently ported from https://github.com/apache/arrow-rs/pull/9848/changes#diff-f2fe5b601534125d0ba033762b62db93ae2f81c71a543b6694df33773ebfa308R939-R976