Parallelize gpuInitStringDescriptors for fixed length byte array data

[mhaseeb123]

Description

Closes #14113

This PR parallelizes the gpuInitStringDescriptors function for the fixed length byte array (FLBA) data at either warp or thread block level via cooperative groups. The function continues to execute serially (thread rank 0 in the group) for variable length arrays.

CC: @etseidl

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[mhaseeb123]

Performance Improvement

Marginal only. Comparison ran several times to and similar improvements seen.

Measured by running Nsight systems on the Gtest ParquetWriterTest.WriteFixedLenByteArray with: constexpr cudf::size_type num_rows = 80'000'000;.

Testbed:

NVIDIA RTX 5880 Ada Generation
AMD Ryzen Threadripper PRO 5975WX 32-Cores
NVIDIA-SMI 550.67
Driver Version: 550.67
CUDA Version: 12.4
Devcontainer: cuda-12.2-pip

Old:

Time (%)  Total Time (ns)  Instances    Avg (ns)       Med (ns)      Min (ns)     Max (ns)    StdDev (ns)                                                   Name   
8.3       34,382,525          1   34,382,525.0   34,382,525.0   34,382,525   34,382,525           0.0  void cudf::io::parquet::detail::<unnamed>::gpuDecodeStringPageData<unsigned char>(cudf::io::parquet…

New:

Time (%)  Total Time (ns)  Instances    Avg (ns)       Med (ns)      Min (ns)     Max (ns)    StdDev (ns)                                                   Name                                                
8.3       34,268,127          1   34,268,127.0   34,268,127.0   34,268,127   34,268,127           0.0  void cudf::io::parquet::detail::<unnamed>::gpuDecodeStringPageData<unsigned char>(cudf::io::parquet…

[vuule]

looks good, few optional suggestions.

[vuule]

would be nice to use tile32 here since we already have it, but I'm not convinced it can be done in a simple way.

[mhaseeb123]

Yes, you are right that we could replace this t < 32 with tile_warp.meta_group_rank() == 0 and it should be good but the logic at L289 is messier to replace since it may be tile_warp.meta_group_rank() == 0 or 1 depending on out_threads0 == 32 or 64 so I left this as is for simplicity.

[vuule]

We could/should probably port the whole logic to thread_groups and avoid the magic 32 multiples. I'd expect that it would not be more complex than the current logic.
Not something for this PR.

[shrshi]

One clarifying question but looks good to me otherwise!

[mhaseeb123]

/merge