Optimize BitArray#reverse!

[HertzDevil]

With BitArray#reverse! available through Indexable::Mutable, this PR adds a much more performant implementation through the addition of the llvm.bitreverse.* intrinsics. These intrinsics are available in all LLVM releases supported by Crystal (latest down to 3.8).

On large BitArrays, this new implementation is ~20x faster for sizes that are multiples of 32, ~13x faster for sizes that aren't. On small BitArrays this becomes a constant-time operation. Benchmarks:

require "benchmark"
require "bit_array"

struct BitArray
  # implementation from `Indexable::Mutable`
  def reverse_old! : self
    return self if size <= 1

    index0 = 0
    index1 = size - 1

    while index0 < index1
      swap(index0, index1)
      index0 += 1
      index1 -= 1
    end

    self
  end
end

module A
  @@ba = uninitialized BitArray
  def self.ba; @@ba; end
  def self.ba=(@@ba); end
end

N = {0, 1, 2, 4, 5, 8, 10, 16, 32, 50, 64, 100, 128, 256, 500, 512, 1000, 1024, 2048, 4096, 5000}

N.each do |n|
  A.ba = BitArray.new(n)
  A.ba.map! { rand > 0.5 }

  Benchmark.ips do |b|
    b.report("old #{n}") do
      A.ba.reverse_old!
    end

    b.report("new #{n}") do
      A.ba.reverse!
    end
  end
end