[FEA] Global thread pool in benchmarks

[vuule]

libcudf currently creates a new thread pool in each multi-threaded benchmark.

This makes reading multithreaded benchmarks very difficult, as the number of unique threads is huge.

Proposed solution:
Files with multithreaded benchmarks should declare a single global pool that's reused in all benchmarks.

[vuule]

reduced the scope to benchmarks since we don't want to invest further in the internal cuFile integration.