RFR: 8302204: Optimize BigDecimal.divide

[Xiaowei Lu]

[JDK-8269667](https://bugs.openjdk.org/browse/JDK-8269667) has uncovered the poor performance of BigDecimal.divide under certain circumstance.

We confront similar situations when benchmarking Spark3 on TPC-DS test kit. According to the flame-graph below, it is StripZeros that spends most of the time of BigDecimal.divide. Hence we propose this patch to optimize stripping zeros.
![图片 1](https://user-images.githubusercontent.com/39413832/218062061-53cd0220-776e-4b72-8b9a-6b0f11707986.png)

Currently, createAndStripZerosToMatchScale() is performed linearly. That is, the target value is parsed from back to front, each time stripping out single ‘0’. To optimize, we can adopt the method of binary search. That is, each time we try to strip out ${scale/2} ‘0’s. 

The performance looks good. Therotically, time complexity of our method is O(log n), while the current one is O(n). In practice, benchmarks on Spark3 show that 1/3 less time (102s->68s) is spent on TPC-DS query4. We also runs Jtreg and JCK to check correctness, and it seems fine.

More about environment: 
we run Spark3.3.0 on Openjdk11, but it seems jdk version doesn’t have much impact on BigDecimal. Spark cluster consists of a main node and 2 core nodes, each has 4cores, 16g memory and 4x500GB storage.

-------------

Commit messages:
 - 8302204: Optimize BigDecimal.divide

Changes: https://git.openjdk.org/jdk/pull/12509/files
 Webrev: https://webrevs.openjdk.org/?repo=jdk&pr=12509&range=00
  Issue: https://bugs.openjdk.org/browse/JDK-8302204
  Stats: 26 lines in 1 file changed: 26 ins; 0 del; 0 mod
  Patch: https://git.openjdk.org/jdk/pull/12509.diff
  Fetch: git fetch https://git.openjdk.org/jdk pull/12509/head:pull/12509

PR: https://git.openjdk.org/jdk/pull/12509