Improve performance for replace-multi for long strings

[davidwendt]

Description

Adds more efficient algorithm for multi-string version of cudf::strings::replace for longer strings (greater than 256 bytes on average in each row).

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]

This code moved to the new multi.cu source file.

[davidwendt]

Rows	Width	Baseline	Change	X
4096	32	0.134	0.114	1.18
4096	64	0.18	0.161	1.12
4096	128	0.291	0.269	1.08
4096	256	0.534	0.51	1.05
4096	512	1.01	0.396	2.55
4096	1024	2.03	0.496	4.09
4096	2048	4.05	0.717	5.65
4096	4096	8.4	1.13	7.43
4096	8192	16.7	2.49	6.71
32768	32	0.135	0.116	1.16
32768	64	0.189	0.172	1.10
32768	128	0.308	0.293	1.05
32768	256	0.565	0.549	1.03
32768	512	1.12	1.09	1.03
32768	1024	2.21	2.08	1.06
32768	2048	4.4	3.45	1.28
32768	4096	8.79	6.03	1.46
32768	8192	17.4	11.1	1.57
262144	32	0.219	0.204	1.07
262144	64	0.409	0.391	1.05
262144	128	0.855	0.834	1.03
262144	256	3.67	3.7	0.99
262144	512	11.9	6.18	1.93
262144	1024	27.8	11.8	2.36
262144	2048	68	23.4	2.91
262144	4096	136	49.1	2.77
2097152	32	0.94	0.91	1.03
2097152	64	2.23	2.2	1.01
2097152	128	5.21	5.12	1.02
2097152	256	15.9	15.7	1.01
2097152	512	52.6	46.8	1.12
16777216	32	6.88	6.81	1.01
16777216	64	17.1	16.9	1.01

Values are in ms and X column is improvement factor. The row-widths 32-256 use the existing code path (no change in performance). For row-widths >256 the new optimized code path is used.

[hyperbolic2346]

Mostly nits here. I like this change and think it will be a great addition!

[PointKernel]

Looks great with one minor non-blocking nit.

Thanks!

[bdice]

Approving CMake changes.

[davidwendt]

/merge