REF: split out grouped masked cummin/max algo

[mzeitlin11]

No user-facing change here, but I think makes logic clearer and future changes easier to understand

Benchmarks look unaffected:

before           after         ratio
[3f67dc33]       [51f8f9c8]
<master>         <masked_cummin_ref>
17.0±2ms         16.3±2ms     0.96  groupby.CumminMax.time_frame_transform('Float64', 'cummax')
17.7±2ms         18.4±2ms     1.04  groupby.CumminMax.time_frame_transform('Float64', 'cummin')
20.8±3ms       16.9±0.3ms    ~0.81  groupby.CumminMax.time_frame_transform('Int64', 'cummax')
20.7±3ms         18.2±3ms    ~0.88  groupby.CumminMax.time_frame_transform('Int64', 'cummin')
15.7±2ms       13.0±0.3ms    ~0.83  groupby.CumminMax.time_frame_transform('float64', 'cummax')
15.5±0.7ms       13.6±0.7ms    ~0.88  groupby.CumminMax.time_frame_transform('float64', 'cummin')
29.8±0.7ms       28.0±0.4ms     0.94  groupby.CumminMax.time_frame_transform('int64', 'cummax')
29.9±1ms         30.0±2ms     1.00  groupby.CumminMax.time_frame_transform('int64', 'cummin')
19.2±2ms         16.9±3ms    ~0.88  groupby.CumminMax.time_frame_transform_many_nulls('Float64', 'cummax')
22.6±2ms         15.6±2ms    ~0.69  groupby.CumminMax.time_frame_transform_many_nulls('Float64', 'cummin')
15.7±3ms         19.3±3ms    ~1.23  groupby.CumminMax.time_frame_transform_many_nulls('Int64', 'cummax')
19.5±2ms         14.7±1ms    ~0.75  groupby.CumminMax.time_frame_transform_many_nulls('Int64', 'cummin')
14.5±2ms       11.4±0.8ms    ~0.79  groupby.CumminMax.time_frame_transform_many_nulls('float64', 'cummax')
13.2±1ms       12.4±0.2ms     0.94  groupby.CumminMax.time_frame_transform_many_nulls('float64', 'cummin')
14.2±1ms       13.1±0.1ms     0.92  groupby.CumminMax.time_frame_transform_many_nulls('int64', 'cummax')
14.1±0.5ms       13.4±0.4ms     0.95  groupby.CumminMax.time_frame_transform_many_nulls('int64', 'cummin')

[jreback]

seems fine. IIRC we have these 2 kernels because the no mask one is >> faster?

[mzeitlin11]

seems fine. IIRC we have these 2 kernels because the no mask one is >> faster?

In an original implementation, I tried forcing a mask (so only use the masked kernel), but there was a slight slowdown for float data because of the isna needed to actually get the mask to pass into the algo.

But all other types where isna is fast (eg nullables, or normal ints) forcing a mask was just as fast. If a small perf hit (I want to say <20%) on float data is acceptable, I can dig up that code and see what the perf hit actually looks like.

Regardless of that choice, I think the split here into 2 clear kernels is helpful because it will make it much simpler if we ever do get rid of one.

[jreback]

thanks @mzeitlin11