feat: Improve benchmark

[ppkarwasz]

This PR:

• Fixes a bug in the benchmark initialization. The decodedData array was not the decoded version of encodedData.
• Tries to figure out the best baseline. I suspect that the simplest version:

  blackhole.consume(new String(decodedData[i].getBytes(UTF_8), UTF_8));

  could be super-optimized by the compiler to something like:

  blackhole.consume(decodedData[i]);

• Adds a benchmark for the toLowerCase() method.

The current benchmark results are:

Benchmark                               (nonAsciiProb)  Mode  Cnt    Score    Error  Units
PercentEncodingBenchmark.baseline                    0  avgt    5   37.273 ±  0.128  us/op
PercentEncodingBenchmark.baseline                  0.1  avgt    5   36.887 ±  0.185  us/op
PercentEncodingBenchmark.baseline                  0.5  avgt    5   36.418 ±  0.548  us/op
PercentEncodingBenchmark.percentDecode               0  avgt    5  158.208 ±  3.255  us/op
PercentEncodingBenchmark.percentDecode             0.1  avgt    5  155.733 ± 29.010  us/op
PercentEncodingBenchmark.percentDecode             0.5  avgt    5  150.590 ± 27.857  us/op
PercentEncodingBenchmark.percentEncode               0  avgt    5  886.282 ± 57.666  us/op
PercentEncodingBenchmark.percentEncode             0.1  avgt    5  879.393 ± 49.545  us/op
PercentEncodingBenchmark.percentEncode             0.5  avgt    5  885.652 ± 57.928  us/op
PercentEncodingBenchmark.toLowerCase                 0  avgt    5  104.868 ±  0.722  us/op
PercentEncodingBenchmark.toLowerCase               0.1  avgt    5  107.073 ±  6.652  us/op
PercentEncodingBenchmark.toLowerCase               0.5  avgt    5  104.674 ±  5.371  us/op
PercentEncodingBenchmark.toLowerCaseJre              0  avgt    5  735.700 ± 25.917  us/op
PercentEncodingBenchmark.toLowerCaseJre            0.1  avgt    5  719.781 ± 36.632  us/op
PercentEncodingBenchmark.toLowerCaseJre            0.5  avgt    5  721.976 ± 31.459  us/op

[jeremylong]

LGTM

[ppkarwasz]

The reason the results for different values of nonAsciiProb look alike is that: there is another bug in the benchmark.

The decodedData and encodedData need to be initialized in a @Setup method, otherwise they'll use the uninitialized value of nonAsciiProbe. 😭 It's incredible how many bugs I could fit in those 4 lines.