Implement datetime columns

[st-pasha]

This is a request to new stypes and functions in support of date/time functionality.

Currently, different packages use different formats for dates and times:

Apache Arrow

• date[day] (32bit, days since the Unix epoch)
• date[ms] (64bit, milliseconds since the Unix epoch)
• time[s] (32bit)
• time[ms] (32bit)
• time[µs] (64bit)
• time[ns] (64bit)
• timestamp[s] (64bit)
• timestamp[ms] (64bit)
• timestamp[µs] (64bit)
• timestamp[ns] (64bit)

When reading from csv, pyarrow parses date/time columns as timestamp[s]. Timestamps with milliseconds are read as strings. Some common date formats like MM/DD/YYYY are not recognized.
The arrow<->pandas conversion guide says that all timestamp[*] formats are converted to pd.Timestamp (np.datetime64[ns]), and all date[*] become object column with datetime.date items.

Pandas

In pandas, there are standalone classes such as pd.Timestamp, pd.Period, pd.DateOffset and pd.Timedelta, but also column-like DatetimeIndex (with dtype datetime64[ns]), and also a Series with dtype datetime64[ns].

When reading from CSV the datetime columns are not parsed, and remain as strings (objects). Conversion can be performed afterwards using to_datetime() function.

Numpy

In numpy all dates are 64-bit, but with various time units: Y, M, W, D, h, m, s, ms, us, ns, ps, fs, as. All datetimes are based on POSIX time with epoch of 1970-01-01T00:00Z.
In numpy 1.6 the default format when parsing was datetime64[us]; since 1.7 the format is selected based on the string.

Python

The datetime module supports classes datetime.date, datetime.time, datetime.datetime and datetime.timedelta. All of these are relatively "heavy" objects, each storing multiple fields: .date has year, month, day; time has hour, min, sec, microsec and timezone; etc.

There is also C-equivalent of datetime module; there the .date object is 4+N bytes, .time is 6+N bytes, and .datetime is 10+N bytes, where N=17, a per-object overhead.

See Also

• https://github.com/HowardHinnant/date - C++14 date library (MIT licensed);
• http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3344.pdf - benchmark of various date implementations;

[XiaomoWu]

IMHO, every option is better than the pandas way. In Rdatatable, date or time is stored as an R object like POSIXct or Date (that's equivalent to the "Python" solution above), or stored internally as an integer like IDate, ITime (that equals to implementing pydatatable's own datetime types).

[st-pasha]

Prerequisite: #1396

[st-pasha]

Done.