Fixes FILL_WITH_MEAN missing value strategy with appropriate cast

[geoffreyangus]

This PR is a follow-up to #2058, which removed the assumption that CSV columns would be loaded with dtypes inferred by Dask/Pandas. Instead, all columns are now initially loaded with dtype object, with the expectation that these are casted to their appropriate dtype downstream. Number features column is therefore not read in with dtype float. This leads to the following error when attempting to apply the missing value strategy FILL_WITH_MEAN:

ludwig/api.py:1276: in preprocess
    preprocessed_data = preprocess_for_training(
ludwig/data/preprocessing.py:1580: in preprocess_for_training
    processed = data_format_processor.preprocess_for_training(
ludwig/data/preprocessing.py:275: in preprocess_for_training
    return _preprocess_file_for_training(
ludwig/data/preprocessing.py:1661: in _preprocess_file_for_training
    data, training_set_metadata = build_dataset(
ludwig/data/preprocessing.py:1099: in build_dataset
    feature_name_to_preprocessing_parameters = build_preprocessing_parameters(
ludwig/data/preprocessing.py:1232: in build_preprocessing_parameters
    fill_value = precompute_fill_value(dataset_cols, feature_config, preprocessing_parameters, backend)
ludwig/data/preprocessing.py:1368: in precompute_fill_value
    return backend.df_engine.compute(dataset_cols[feature[COLUMN]].mean())
venv38/lib/python3.8/site-packages/dask/dataframe/core.py:94: in wrapper
    return func(self, *args, **kwargs)
venv38/lib/python3.8/site-packages/dask/dataframe/core.py:1986: in mean
    _raise_if_object_series(self, "mean")
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

x = Dask Series Structure:
npartitions=1
    object
       ...
Name: num_E23AE, dtype: object
Dask Name: getitem, 2 tasks, funcname = 'mean'

    def _raise_if_object_series(x, funcname):
        """
        Utility function to raise an error if an object column does not support
        a certain operation like `mean`.
        """
        if isinstance(x, Series) and hasattr(x, "dtype") and x.dtype == object:
>           raise ValueError("`%s` not supported with object series" % funcname)
E           ValueError: `mean` not supported with object series

venv38/lib/python3.8/site-packages/dask/dataframe/core.py:3161: ValueError

This PR fixes this with appropriate dtype casting right before calling mean(). This PR also adds a regression test.

[github-actions]

Unit Test Results

       4 files  ±0         4 suites  ±0   1h 20m 27s ⏱️ -40s
2 825 tests +2  2 791 ✔️ +2  34 💤 ±0  0 ❌ ±0 
5 650 runs  +4  5 579 ✔️ +4  71 💤 ±0  0 ❌ ±0 

Results for commit 799fdc5. ± Comparison against base commit 520af82.

[justinxzhao]

LGTM. Nice parametrization of the backend.