Extending the grouper `base` argument

[LucaAmerio]

Problem description

pd.Grouper's base argument is almost undocumented (link). Despite this, the argument is extremely useful when grouping a dataframe with uneven sampling rate with an arbitrary start time.

Take for example the case of a dataframe representing a 60 minutes experiment with uneven sampling rate starting at '12-01-2019 11:55:23.01938' that we want to split in six 10-minutes group. base is our best ally here.

Despite this, in the current implementation base represents a floating number of minutes. It's however quite hard to understand what this number refers to, is quite cumbersome to convert the above '11:55:23.01938' into a number of minutes and gives you little flexibility regarding when the grouping will start.

This can also cause rounding error issues as descibed here.

I suggest replacing (or extend) this allowing it to be a datetime object so that the edge of each group is simply
base + N * freq
where N is an integer number. This is much cleaner and easy to understand.

[gfyoung]

cc @jreback regarding the enhancement

@pj1989 : We'll probably need to discuss the enhancement further, but in the meantime, documentation on base would definitely be appreciated!