Grouping by frequency and resampling

[shwina]

Closes #6255, #8416

This PR implements two related features:

1. Grouping by a frequency via the freq= argument to cudf.Grouper
2. and time-series resampling via the .resample() API

Either operation results in _Resampler object that represents the data resampled into "bins" of a particular frequency. The following operations are supported on resampled data:

1. Aggregations such as min() and max(), performed bin-wise
2. ffill() and bfill() methods: forward and backward filling in the case of upsampling data
3. asfreq(): returns the resampled data as a Series or DataFrame()

These are all best understood by example:

First, we create a time series with 1 minute intervals:

>>> index = cudf.date_range(start="2001-01-01", periods=10, freq="1T")
>>> sr = cudf.Series(range(10), index=index)
>>> sr
2001-01-01 00:00:00    0
2001-01-01 00:01:00    1
2001-01-01 00:02:00    2
2001-01-01 00:03:00    3
2001-01-01 00:04:00    4
2001-01-01 00:05:00    5
2001-01-01 00:06:00    6
2001-01-01 00:07:00    7
2001-01-01 00:08:00    8
2001-01-01 00:09:00    9
dtype: int64

Downsampling to 3 minute intervals, followed by a "sum" aggregation:

>>> sr.resample("3T").sum()  # equivalently, sr.groupby(cudf.Grouper(freq="3T")).sum()
2001-01-01 00:00:00     3
2001-01-01 00:03:00    12
2001-01-01 00:06:00    21
2001-01-01 00:09:00     9
dtype: int64

Upsampling to 30 second intervals:

>>> sr.resample("30s").asfreq()
2001-01-01 00:00:00    0.0
2001-01-01 00:00:30    NaN
2001-01-01 00:01:00    1.0
2001-01-01 00:01:30    NaN
2001-01-01 00:02:00    2.0
2001-01-01 00:02:30    NaN
2001-01-01 00:03:00    3.0
2001-01-01 00:03:30    NaN
2001-01-01 00:04:00    4.0
2001-01-01 00:04:30    NaN
2001-01-01 00:05:00    5.0
2001-01-01 00:05:30    NaN
2001-01-01 00:06:00    6.0
2001-01-01 00:06:30    NaN
2001-01-01 00:07:00    7.0
2001-01-01 00:07:30    NaN
2001-01-01 00:08:00    8.0
2001-01-01 00:08:30    NaN
2001-01-01 00:09:00    9.0
Freq: 30S, dtype: float64

Upsampling to 30 second intervals, followed by a forward fill:

>>> sr.resample("30s").ffill()
2001-01-01 00:00:00    0
2001-01-01 00:00:30    0
2001-01-01 00:01:00    1
2001-01-01 00:01:30    1
2001-01-01 00:02:00    2
2001-01-01 00:02:30    2
2001-01-01 00:03:00    3
2001-01-01 00:03:30    3
2001-01-01 00:04:00    4
2001-01-01 00:04:30    4
2001-01-01 00:05:00    5
2001-01-01 00:05:30    5
2001-01-01 00:06:00    6
2001-01-01 00:06:30    6
2001-01-01 00:07:00    7
2001-01-01 00:07:30    7
2001-01-01 00:08:00    8
2001-01-01 00:08:30    8
2001-01-01 00:09:00    9
Freq: 30S, dtype: int64

[shwina]

rerun tests

[isVoid]

Currently when I run it with any operation, because IndexedFrame.sort_values calls take, I get a FutureWarning which is quite confusing to user:

In [10]: gdf.groupby(cudf.Grouper(key="Publish date", freq="1h", label="left", closed="left")).sum()
/raid/wangm/dev/rapids/cudf/python/cudf/cudf/core/frame.py:3076: FutureWarning: keep_index is deprecated and will be removed in the future.
  FutureWarning,
Out[10]: 
                       ID Price
Publish date                   
2000-01-01 12:00:00     1    30
2000-01-01 13:00:00  <NA>  <NA>
2000-01-01 14:00:00  <NA>  <NA>
...

This issue is not specific to this PR - raising an issue to track it instead. #9667

[vyasr]

This could all be a single conditional

[vyasr]

I deleted a few comments of the comments I made here. I already planned to do some refactoring of cut, and reading this I see quite a bit more that should be done, but it's out of scope for this PR. No need to change anything, we can revisit it at a later point.

[quasiben]

@gpucibot merge

[shwina]

@gpucibot merge