Consider using of tzcnt instruction for better performance in foreach loop and iterator of bitsets

[plokhotnyuk]

This instruction has a low latency and throughput on most contemporary CPUs.

Also it is implemented as intrinsic and turned on by default in JDK 8:

$ java -XX:+PrintFlagsFinal -version | grep Trailing
     bool UseCountTrailingZerosInstruction          = true                                {ARCH product}
java version "1.8.0_192"
Java(TM) SE Runtime Environment (build 1.8.0_192-b12)
Java HotSpot(TM) 64-Bit Server VM (build 25.192-b12, mixed mode)

Below is a snippet how it can be used:

  override def foreach[U](f: Int => U): Unit = {
    val n = nwords
    var i = 0
    while (i < n) {
      var w = word(i)
      while (w != 0) {
        f(java.lang.Long.numberOfTrailingZeros(w) + (i << 6))
        w &= w - 1
      }
      i += 1
    }
  }

Inspired by the Daniel Lemire's blog post.

[linasm]

@plokhotnyuk I am really seeing a nice improvement for BitSet.iterator. PR on its way.
foreach is a bit complicated though - there is an improvement for sparse BitSets, but degradation for dense ones, I guess it needs more investigation.