remove pyarrow dependency for polars

[jmoralez]

• The CV methods used pl.from_pandas, which requires pyarrow to be installed. This moves the processing to the corresponding engine so that the dataframe is built from the start in pandas/polars, thus removing the need to convert a pandas dataframe to polars and the need for pyarrow.
• Also vectorizes the times creation, which should speed up all methods.
• Adds the ufp prefix to the utilsforecast.processing functions.
• Only saves the original times because the ids can be easily reconstructed when computing the fitted values, which saves memory.
• Fixes predict and forecast with polars dataframe, output ds has different time unit than source #703.
• Deletes the fcst_fitted_values_ and cv_fitted_values_ attributes at the start of the forecast and cross_validation methods to avoid getting old results back, e.g.

sf.forecast(df=df, ..., fitted=True)
sf.forecast(df=other_df, ..., fitted=False)
sf.forecast_fitted_values()  # these would correspond to the first df. in this PR it raises an error

Breaking changes

• The frequency when using polars dataframes must be a valid by argument for offset_by, e.g. daily frequency is '1d'.
• The frequency when using integer times should be an integer, e.g. if you have ds = [1, 2, 3] you should set freq=1

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[AzulGarza]

really nice the abstraction of process_df and cv_times in utilsforecast😍