First check for BaseDtype when infering the data type of an arbitrary object

[shwina]

We have an internal utility called dtype() that attempts to infer the data type of an arbitrary object. One of the first thing that dtype() does is attempt to call np.dtype(obj). That can be slow for extremely large cardinality categorical data types, as it copies data to host (in particular, it attempts to call the object's __repr__):

Before this PR:

dtype = cudf.CategoricalDtype(categories=range(100_000_000))
%%time x = cudf.core.dtypes.dtype(dtype)
CPU times: user 3.75 s, sys: 885 ms, total: 4.64 s
Wall time: 4.63 s

This PR ensures we attempt to do far less expensive inference first, before calling np.dtype(...).

After this PR:

%%time x = cudf.core.dtypes.dtype(dtype)
CPU times: user 13 µs, sys: 1 µs, total: 14 µs
Wall time: 19.1 µs

[bdice]

Good catch. How did you come across this?

[wence-]

Ooof.

[shwina]

/merge