[FEA] Add support for `datetime32[D]` in cuDF python

[GregoryKimball]

Is your feature request related to a problem? Please describe. Libcudf support 5 timestamp types, but cuDF python only supports 4.

libcudf type name	libcudf physical type	cuDF type	Arrow type
TIMESTAMP_DAYS	int32	n/a	Date32Type
TIMESTAMP_SECONDS	int64	datetime64[s]	TimeUnit::SECOND
TIMESTAMP_MILLISECONDS	int64	datetime64[ms]	TimeUnit::MILLI
TIMESTAMP_MICROSECONDS	int64	datetime64[us]	TimeUnit::MICRO
TIMESTAMP_NANOSECONDS	int64	datetime64[ns]	TimeUnit::NANO

Describe the solution you'd like Adjust cuDF python type support to leverage all of the available libcudf types.

Additional context I was generating TPC-H files for Velox benchmarking, and cuDF python cannot write files with int32 timestamp days type. This means that Velox fails to execute scalar functions without additional query changes to cast from Velox TIMESTAMP to Velox DATE.

Error:
 terminate called after throwing an instance of 'facebook::velox::VeloxUserError'
  what():  Exception: VeloxUserError
Error Source: USER
Error Code: INVALID_ARGUMENT
Reason: Scalar function signature is not supported: between(TIMESTAMP, DATE, DATE). Supported signatures: (timestamp,timestamp,timestamp) -> boolean, (date,date,date) -> boolean, (real,real,real) -> boolean, (double,double,double) -> boolean, (bigint,bigint,bigint) -> boolean, (decimal(i1,i5),decimal(i1,i5),decimal(i1,i5)) -> boolean, (varchar,varchar,varchar) -> boolean, (integer,integer,integer) -> boolean, (smallint,smallint,smallint) -> boolean, (tinyint,tinyint,tinyint) -> boolean.
Retriable: False
Function: resolveScalarFunctionType
File: /nfs/repo/velox2/velox/parse/TypeResolver.cpp
Line: 99

[mroeschke]

Just noting that pandas does not support datetime32/64[D] and I don't think has plans to (because daily resolution logic gets murky with timezone data and DST), but should still be feasible for cudf Python

[vyasr]

That suggests to me that we want this support in pylibcudf, but not in cuDF Python. That wouldn't be as convenient for the original use case that @GregoryKimball was interested in though.