[FEA] Parallelize gpuInitStringDescriptors when Parquet input type is FIXED_LEN_BYTE_ARRAY

[etseidl]

As part of the preprocessing of PLAIN encoded string data in the parquet reader, a pass through the page data is performed to either gather string sizes, or initialize {ptr, length} tuples for use by the decoder. For variable width string data, this pass must be performed by a single thread. But in the case of fixed width data, all threads in the warp should be able to participate.

cudf/cpp/src/io/parquet/page_decode.cuh

Line 409 in 1bfeee7

__device__ size_type gpuInitStringDescriptors(page_state_s volatile* s,