support nanosecond precision values in TimeDtype

[tswast]

datetime.time only supports microsecond precision.

Similarly, the BigQuery TIME data type only returns microsecond precision.

The pandas convention is to use nanosecond precision for timestamps via the pandas.Timestamp class. It may make sense for this extension dtype to also support this.

Subtasks:

• Preserve nanoseconds when we are given them in the data (e.g. from numpy array or pyarrow array).
• Add unit as an optional argument in the dtype constructor, similar to https://github.com/pandas-dev/pandas/blob/1c986d6213904fd7d9acc5622dc91d029d3f1218/pandas/core/dtypes/dtypes.py#L777 -- Note: maybe unnecessary since we always store as nanoseconds, but could be useful as a box/unbox hint (see below)
• Return / accept something other than datetime.time that supports nanoseconds instead of only microsecond precision. This is similar to datetime.datetime versus pandas.Timestamp.

[jbrockmendel]

Could wrap TimedeltaArray and disallow anything not between 0 and 24H

[tswast]

#39 makes some tiny progress towards this. If going from arrow -> pandas -> arrow, nanosecond-precision times should be preserved.

[tswast]

Could wrap TimedeltaArray and disallow anything not between 0 and 24H

That's more-or-less what we're doing already. We store the right precision data, but when folks access individual rows, they get back a datetime.time which doesn't have a nanosecond field. Need something similar to pandas.Timestamp.

[chalmerlowe]

Closing as will not work. Efforts going forward are focused on more fully interacting with ArrowDtype.