Branch-less alphanumeric underscore's replacement

[franz1981]

This is following the same idea of quarkusio/quarkus#45546 re using lookup tables.
If the src Strings are rarely latin only beyond US_ASCII i.e. [128, 255] this can be further simplified by having a 128-sized lookup table.

I didn't yet looked at its performance - and, as suggested at mkouba/qute-benchmarks#1 we should have a proper micro-benchmark which account to stress the branch-predictor too (or not - should be configurable) - unless the non-alphanumeric case is supposed to be very predictable in reality as well

[franz1981]

I probably have a further suggestion here

[franz1981]

I've created this benchmark at which gives these numbers franz1981/java-puzzles@8fb833c

with highly predictable sameples:

Benchmark                                                (finalDoubleQuoteProbability)  (nonAlphanumericProbability)  (samples)  (size)  Mode  Cnt   Score   Error  Units
ReplaceWithUnderscores.replaceWithUnderscoresSwitch                                 10                            10        100      33  avgt   10  94.525 ± 3.802  ns/op
ReplaceWithUnderscores.replaceWithUnderscoresTableArray                             10                            10        100      33  avgt   10  20.951 ± 0.339  ns/op
ReplaceWithUnderscores.replaceWithUnderscoresTableSb                                10                            10        100      33  avgt   10  77.077 ± 1.501  ns/op

and with not that much predictable (10K samples):

Benchmark                                                     (finalDoubleQuoteProbability)  (nonAlphanumericProbability)  (samples)  (size)  Mode  Cnt    Score   Error  Units
ReplaceWithUnderscores.replaceWithUnderscoresSwitch                                      10                            10      10000      33  avgt   10  231.371 ± 9.276  ns/op
ReplaceWithUnderscores.replaceWithUnderscoresTableArray                                  10                            10      10000      33  avgt   10   32.910 ± 2.189  ns/op
ReplaceWithUnderscores.replaceWithUnderscoresTableSb                                     10                            10      10000      33  avgt   10  102.949 ± 4.079  ns/op

In both case the Table version (same of this PR) is faster with the TableSb (using the builder) to be slower .

I didn't accounted yet for the code duplication, but since native image is not wow to handle StringBuilders and Strings - moving to a "ligher" abstractions should pay off (@galderz can verify it :P) - for JIT I would like to test when only C1 is running or with "cold" code as well.

[franz1981]

Sadly if I got right what C1MaxInlineSize does - C1 doesn't have inlining superpowers, hence reusing code could be relatively costly if the callee is not trivially inlinable, which seems this case