Add nanquantile function

[max-sixty]

An attempt to solve #161

But:

• I can't get it to broadcast with guvectorize and take multiple quantiles. Somewhat understandably, it doesn't know that quantiles is a fixed list of values, rather than an array to broadcast over.
• It doesn't seem any faster! (forgive the bad output, scroll right for timings — it's the same for all arrays apart from small arrays, where it's much slower)

I haven't done benchmarks for the version at https://krstn.eu/np.nanpercentile()-there-has-to-be-a-faster-way/, or xclim's version.

If anyone wants to take this and test if those versions are indeed faster, that would be helpful to see.

                       ok
[33.33%] ··· ========================================================================================================================= ======== ==========
                                                                        func                                                              n
             ------------------------------------------------------------------------------------------------------------------------- -------- ----------
              (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)     10     246±0μs
              (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)    1000    4.79±0ms
              (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)   100000   453±0ms
             ========================================================================================================================= ======== ==========

[66.67%] ··· benchmarks.Funcs.time_numpy                                                                                                                                       ok
[66.67%] ··· ========================================================================================================================= ======== ==========
                                                                        func                                                              n
             ------------------------------------------------------------------------------------------------------------------------- -------- ----------
              (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)     10     55.9±0μs
              (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)    1000    4.71±0ms
              (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)   100000   475±0ms
             ========================================================================================================================= ======== ==========

[100.00%] ··· benchmarks.Funcs.time_pandas                                                                                                                                      ok
[100.00%] ··· ========================================================================================================================= ======== ==========
                                                                         func                                                              n
              ------------------------------------------------------------------------------------------------------------------------- -------- ----------
               (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)     10     128±0μs
               (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)    1000    4.74±0ms
               (<numba._GUFunc 'nanquantile_single'>, <function Funcs.<lambda> at 0x15d24a310>, <function nanquantile at 0x1026a0870>)   100000   448±0ms
              ========================================================================================================================= ======== ==========

[max-sixty]

OK — actually this is better news — np.nanquantile is only slow when it's parallelizing with e.g. axis=0, because the slow part is how it vectorizes.

So the current benchmarks have, for a 300x100000 array:

• numbagg.nanquantile at 578ms
• np.nanquantile at 2800ms
• pandas .quantile function at 451ms — faster than both

So adding this to numbagg probably does make sense. Unfortunately we'll probably only be able to run for a single quantile, as long as we stay using gufuncs, which allow this over any number of dimensions. Though very open to ideas for a way of fixing that...

I need to clean up the code a lot.

[max-sixty]

OK, actually a couple of hours of work showed me this is very possible, by managing the axes parameter to a gufunc.

So this works, and is about 4x faster than np.nanquantile. It's about 10% slower than np.quantile. pandas is surprisingly fast here — we're ~20% slower than its .quantile. (Though pandas' .quantile doesn't have the benefits of a gufunc, such as vectorizing over any combination of axes.) I think that's probably because numbagg is less fast for sorting, whereas for standard loops, it's extremely fast.

I need to clean up the benchmarks, but I'll merge for now.

This has mostly been me, but if anyone has substantive feedback on the code, that would be very welcome.