PERF: avoid object conversion in fillna(method=pad|backfill) for masked arrays

[simonjayhawkins]

[ 75.00%] ··· frame_methods.Fillna.time_frame_fillna                                                                         ok
[ 75.00%] ··· ========= ======== ============ ============= ========== ============= =============
              --                                               dtype                              
              ------------------ -----------------------------------------------------------------
               inplace   method    float64       float32      object       Int64        Float64   
              ========= ======== ============ ============= ========== ============= =============
                 True     pad     3.44±0.3ms    2.64±0.2ms   307±10ms    3.97±0.2ms    3.72±0.1ms 
                 True    bfill    3.23±0.2ms    2.89±0.2ms   275±8ms     2.34±0.1ms   2.45±0.09ms 
                False     pad     3.62±0.1ms   2.06±0.01ms   181±3ms    4.23±0.07ms   4.16±0.07ms 
                False    bfill    3.59±0.2ms   2.06±0.02ms   186±6ms    4.35±0.09ms   4.18±0.06ms 
              ========= ======== ============ ============= ========== ============= =============

       before           after         ratio
     [2e5c28fc]       [d3ae0cf1]
     <master>         <BaseMaskedArray.fillna>
-     2.28±0.06ms      2.06±0.02ms     0.91  frame_methods.Fillna.time_frame_fillna(False, 'bfill', 'float32')
-      15.5±0.2ms      2.45±0.09ms     0.16  frame_methods.Fillna.time_frame_fillna(True, 'bfill', 'Float64')
-      16.5±0.2ms       2.34±0.1ms     0.14  frame_methods.Fillna.time_frame_fillna(True, 'bfill', 'Int64')
-      85.2±0.9ms      4.18±0.06ms     0.05  frame_methods.Fillna.time_frame_fillna(False, 'bfill', 'Float64')
-      86.6±0.6ms      4.16±0.07ms     0.05  frame_methods.Fillna.time_frame_fillna(False, 'pad', 'Float64')
-      90.9±0.8ms      4.35±0.09ms     0.05  frame_methods.Fillna.time_frame_fillna(False, 'bfill', 'Int64')
-      92.4±0.4ms      4.23±0.07ms     0.05  frame_methods.Fillna.time_frame_fillna(False, 'pad', 'Int64')
-        85.3±1ms       3.72±0.1ms     0.04  frame_methods.Fillna.time_frame_fillna(True, 'pad', 'Float64')
-        92.5±2ms       3.97±0.2ms     0.04  frame_methods.Fillna.time_frame_fillna(True, 'pad', 'Int64')

SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.
PERFORMANCE INCREASED.

[simonjayhawkins]

ExtensionBlock.interpolate creates another copy L1774. if we always copy in EA.fillna then that copy shouldn't be needed.

[jreback]

rather than adding new, can we simply always use the mask (for all dtypes)

[simonjayhawkins]

probably, the 1d versions of pad and backfill are only used by EA.fillna and Series.replace.

some of the EAs override fillna and use 2d algorithms. NDArrayBackedExtensionArray uses the 1d versions but should be using a 2d version (not an issue since only called columnwise, but would be fail if called directly on a 2d EA) The other EAs that use this convert to object so are not performant anyway.

not sure about Series.replace, will look some more.

[jreback]

we don't support any 2D EA atm, so can leave those out for now

[simonjayhawkins]

done. if we are always updating and returning the mask, maybe we should return the same dtype as the original mask, or only allow boolean mask in python space and always return a boolean mask.

[jreback]

looks good. pls add a whatsnew note. you had a comment about multiple copies ; can try to handle here or make a new issue. ping when ready.

[jreback]

thanks @simonjayhawkins very nice

[jorisvandenbossche]

This was failing CI (now also on master):

=================================== FAILURES ===================================
________________ TestMissing.test_fillna_no_op_returns_copy[0] _________________
[gw1] linux -- Python 3.8.8 /home/vsts/miniconda3/envs/pandas-dev/bin/python

self = <pandas.tests.extension.test_sparse.TestMissing object at 0x7f0590718cd0>
data = [7, 33, 0, 95, 38, 0, 30, 26, 0, 47, 6, 0, 66, 4, 0, 57, 94, 0, 19, 23, 0, 48, 3, 0, 75, 74, 0, 51, 74, 0, 13, 73, 0, ...66, 67, 69, 70, 72, 73, 75,
       76, 78, 79, 81, 82, 84, 85, 87, 88, 90, 91, 93, 94, 96, 97, 99],
      dtype=int32)

request = <FixtureRequest for <Function test_fillna_no_op_returns_copy[0]>>

    def test_fillna_no_op_returns_copy(self, data, request):
        if np.isnan(data.fill_value):
            request.node.add_marker(
                pytest.mark.xfail(reason="returns array with different fill value")
            )
>       super().test_fillna_no_op_returns_copy(data)

pandas/tests/extension/test_sparse.py:229: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
pandas/tests/extension/base/missing.py:80: in test_fillna_no_op_returns_copy
    result = data.fillna(method="backfill")
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = [7, 33, 0, 95, 38, 0, 30, 26, 0, 47, 6, 0, 66, 4, 0, 57, 94, 0, 19, 23, 0, 48, 3, 0, 75, 74, 0, 51, 74, 0, 13, 73, 0, ...66, 67, 69, 70, 72, 73, 75,
       76, 78, 79, 81, 82, 84, 85, 87, 88, 90, 91, 93, 94, 96, 97, 99],
      dtype=int32)

value = None, method = 'backfill', limit = None

    def fillna(self, value=None, method=None, limit=None):
        """
        Fill missing values with `value`.
>           warnings.warn(msg, PerformanceWarning)
E           pandas.errors.PerformanceWarning: fillna with 'method' requires high memory usage.

pandas/core/arrays/sparse/array.py:672: PerformanceWarning