[Torchscript] Add Date feature preprocessing

[geoffreyangus]

This PR implements Date feature preprocessing in Torchscript.

Date feature preprocessing in Torchscript deviates from vanilla preprocessing in one key way: instead of being able to pass in the raw date-formatted string into the torchscript model, the user must first featurize it into a vector with the values found in vector (below).

yearday = datetime_obj.toordinal() - date(datetime_obj.year, 1, 1).toordinal() + 1

midnight = datetime_obj.replace(hour=0, minute=0, second=0, microsecond=0)
second_of_day = (datetime_obj - midnight).seconds

vector = [
    datetime_obj.year,
    datetime_obj.month,
    datetime_obj.day,
    datetime_obj.weekday(),
    yearday,
    datetime_obj.hour,
    datetime_obj.minute,
    datetime_obj.second,
    second_of_day,
]

A utility function for converting datetime objects into this exact type of vector can be found at ludwig.features.date_feature.create_vector_from_datetime_obj.

This deviation occurs because the datetime and dateutil libraries are not currently scriptable (https://pytorch.org/docs/stable/jit.html#python-functions-and-modules).

In general, if the user plans on deploying with Torchscript with a Python backend, they can use the utility function ludwig.models.inference.to_inference_module_input_from_dataframe to ensure that their input features are being passed into the torchscript model correctly.

[github-actions]

Unit Test Results

       6 files  ±0         6 suites  ±0   2h 20m 1s ⏱️ + 20m 22s
2 831 tests +1  2 797 ✔️ +1    34 💤 ±0  0 ❌ ±0 
8 493 runs  +3  8 387 ✔️ +3  106 💤 ±0  0 ❌ ±0 

Results for commit 9abc608. ± Comparison against base commit fdcbcc7.