Support Ascii function for ascii and latin-1 [databricks]

[thirtiseven]

Closes #9585

This PR part supported ascii function for strings starting with ASCII and Latin-1 supplement, returning results from 0 to 255. The function is disabled by default.

This PR uses cudf::code_points to get the utf-8 code points of the first letter in a string, and converts to ASCII using the following rules:

ascii	utf-8 dec	workaround
0~127	0~127	cudf::code_points
128~191	49792~49855	code_points - 49,664
192~255	50048~50111	code_points -49,856

It is good enough for customer, we can file another issue to fully support it when needed.

perf test results

data: 50000000 string from big datagen:

spark.time(df.selectExpr("COUNT(ascii(a)) as asc0", "COUNT(ascii(a)) as asc1", "COUNT(ascii(a)) as asc2", "COUNT(ascii(a)) as asc3", "COUNT(ascii(a)) as asc4", "COUNT(ascii(a)) as asc5", "COUNT(ascii(a)) as asc6", "COUNT(ascii(a)) as asc7", "COUNT(ascii(a)) as asc8", "COUNT(ascii(a)) as asc9").show())

GPU Time (ms)	CPU Time (ms)	Speed up
1,653	2,787	69%

[thirtiseven]

Did some perf tests, and gpu is slightly faster than cpu (12%). I plan to refine my implementation to reduce times of cuDF API calling, but the undefined values will be more undefined, which doesn't matter for now.

[thirtiseven]

Ok, I refined the implementation and it now runs faster, about 69% speedup over cpu.

[thirtiseven]

build

[thirtiseven]

Test failed in CI because Latin-1 support in Spark is from 3.3.1, will add shim for it so ascii under 3.3.1 will be faster and (maybe) fully supported.

[thirtiseven]

build

[revans2]

Just need to be sure that we test this on databricks

[thirtiseven]

build

[thirtiseven]

build

[revans2]

Looks like we need to fix this for 3.5.1 too

[thirtiseven]

build

[thirtiseven]

perf test results on the byte-based solution (in lower version shims):

GPU Time (ms)	CPU Time (ms)	Speed up
541.4	2,020.6	3.73x

[revans2]

The previous tests timed out trying to run on databricks. Trying again.

[revans2]

build