Performance improvement in cudf::strings::find/rfind for long strings

[davidwendt]

Description

Improves performance for longer strings with cudf::strings::find() and cudf::strings::rfind() APIs.

The current implementation works well with small-ish strings and so this new implementation splits into a long-ish string algorithm when the average number of bytes per string is 64 bytes or greater. The new implementation searches for the target string by applying a warp per string kernel. Additionally, the special logic needed for matching an empty target (easily checked in host code) is factored out into its own transform functor.
For longer strings, the performance improvement is about 2-6x.

Reference #13048

Checklist

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

[davidwendt]

Some benchmark results from benchmarks/STRINGS_BENCH --benchmark_filter=StringFindScalar/find/

rows	width	baseline	warp	x
4096	32	0.037	0.037	1.00
4096	64	0.045	0.045	1.00
4096	128	0.062	0.062	1.00
4096	256	0.102	0.040	2.55
4096	512	0.169	0.048	3.52
4096	1024	0.292	0.061	4.79
4096	2048	0.463	0.084	5.51
4096	4096	0.802	0.141	5.69
4096	8192	1.40	0.238	5.88
32768	32	0.040	0.040	1.00
32768	64	0.048	0.047	1.02
32768	128	0.065	0.065	1.00
32768	256	0.106	0.074	1.43
32768	512	0.197	0.108	1.82
32768	1024	0.326	0.153	2.13
32768	2048	0.524	0.240	2.18
32768	4096	0.890	0.415	2.14
32768	8192	1.48	0.766	1.93
262144	32	0.061	0.061	1.00
262144	64	0.091	0.091	1.00
262144	128	0.177	0.177	1.00
262144	256	0.788	0.339	2.32
262144	512	1.40	0.558	2.51
262144	1024	2.08	0.882	2.36
262144	2048	5.75	1.45	3.97
262144	4096	15.3	2.69	5.69
2097152	32	0.215	0.214	1.00
2097152	64	0.403	0.402	1.00
2097152	128	0.941	0.941	1.00
2097152	256	6.40	2.45	2.61
2097152	512	9.94	4.16	2.39
16777216	32	1.44	1.43	1.01
16777216	64	2.87	2.87	1.00

[PointKernel]

Nice work! 🔥

[mythrocks]

Sorry for the delay. It took me ages to understand the new kernel.

[davidwendt]

/merge