[v10.5.2] Performance Improvement of PER

[ymd-h]

At version 10.5.2, we improved the internal implementation of Segment Tree for PER. (It was one of the oldest section.)

On my local machine, Sum Tree with size of 10^6 improved;

1. Multi-step insertion: 1.5 - 2 times faster
   • Surprisingly, it was slower than multiple single-step insertion, but is faster now.
2. Index finding: 20 times faster
   • It was comparable with No.3, but is trivial now.
3. Full update of internal partial sum structure: 1.5 - 2 times faster
   • This is still one of the heaviest operation for multiprocessing version.

These improvements also affect PrioritizedReplayBuffer and MPPrioritizedReplayBuffer to some extent.

We continue to investigate further performance improvement.