Make TensorBoard histogram processing 3x faster

We discovered that on one of our datasets, the TensorBoard backend was
spending 80% of its time processing histograms. Particularly the
histogram compression routine, which did a lot of math in pure Python
on raw histograms which are potentially quite large considering the C++
Histogram class uses a default bucket count of 774.

These routines have now been optimized so that the _ProcessHistogram
method of EventAccumulator and its subroutines take 39% of the time
they did previously. This has been accomplished by doing the following:

- Use numpy fast vector math to compute cumulative sum
- Use numpy to binary search compression points in cumulative sum
- Only convert histo.bucket to a list once rather than twice
Change: 132285757