Depends on #6000

Current cuIO benchmarks only use random data with low repetition. Encode/decode of some formats (ORC, Parquet...) varies significantly depending on the data profile.
Also, newly supported data types (like lists) are not covered.
As of now, the missing cases are:

• Reading RLE encoded ORC files.
• Reading RLE encoded Parquet files.
• Writing RLE-friendly data t

Right now the macros for banded alignment like CUDAPOA_BANDED_MATRIX_RIGHT_PADDING are not shared between the kernel code and other sources that determine sizes for bands (such as in batch.cu). For now they're hard coded to specific numbers, but this is error prone. This needs to be fixed to the macros are shared between files.

When running the regression, the resulting logs seem to end up in third_party/tests/$TEST/.... This of course is not unnoticed by git, so a git status shows a ton of non-added new files added.

To reproduce:

make
make regression
git status   # observe all the files

Build or test artifacts should never clutter the rest of the code-base (we should regard them as read-only in

/ɛ, ɔ/ appear in a few-dozen phonemic transcriptions, but they either are allophones of, or incorrect transcriptions, of, /e, o/ in the modern language.