[NSE-857] Further optimizations of validity buffer split

[FelixYBW]

What changes were proposed in this pull request?

Previous implementation of validity buffer split performance is too bad.

no validity buffer	1.55
w/ validity buffer, before opt	2.51
w/ validity buffer, after opt	2.01

From profile data, validity buffer split still takes more time than value copy. Most of the time is spending on the first touch of dst buffer

67.32%  BenchmarkShuffl  libspark_columnar_jni.so  [.] sparkcolumnarplugin::shuffle::Splitter::SplitBoolType                                                                              
25.96%  BenchmarkShuffl  libspark_columnar_jni.so  [.] sparkcolumnarplugin::shuffle::Splitter::SplitFixedWidthValueBuffer                                                                 
 1.81%  BenchmarkShuffl  libspark_columnar_jni.so  [.] sparkcolumnarplugin::shuffle::RoundRobinSplitter::ComputeAndCountPartitionId                                                       
 1.48%  BenchmarkShuffl  libc-2.17.so              [.] __memmove_ssse3_back                                                                                                               
 1.31%  BenchmarkShuffl  libspark_columnar_jni.so  [.] sparkcolumnarplugin::shuffle::Splitter::DoSplit                                                                                    
 1.02%  BenchmarkShuffl  libspark_columnar_jni.so  [.] sparkcolumnarplugin::shuffle::Splitter::SplitFixedWidthValidityBuffer    

How was this patch tested?

[github-actions]

#857