bugfix/granularity handling fix

[elineii]

#1

• Fixed a bug with incorrect determination of common types of frequency of timestamps in time series.
• Improved automatic granularity expansion.
• Added possibility to add custom granularity settings (new part in tutorial).

If the rows do not contain segments that are too short (that are shorter than history + horizon), then tsururu will try to extract the row granularity on its own. Сurrently the following types are supported:

• Yearly (and YearlyEnd)
• Quarterly (and Quarterly)
• Monthly (and MonthlyEnd)
• Weekly
• Daily
• Hourly
• Minlutely
• Secondly
• Microsecondly

There is also support for compound granularities (10 days, 15 minutes, 32 seconds, etc.).

It is possible to set your own granularity using the pd.DateOffset class or related classes from pandas.tseries.offsets, which must be fed as delta parameter into the Dataset class.

#5

Fixed a bug with piplane crash if features have specific characters in their names (e.g. brackets) and if features have nested names (e.g. feature, feature0, feature1, etc.).

Fixed bug with drop_raw_feature=True, now this parameter works as intuitively expected

Fixed a bug with predictions changing when the order of transformer application passing is changed (more specifically, features are now fed to the ML model in lexicographical order).

Fixed situation with need to sort dataset by id and date before initialization of TSDataset.

[DESimakov]

Merge with TODO refactor feature pipeline with new logic