Refactor background classes

[larrybradley]

This PR is a refactor of the background classes, including Background2D, motivated to improve their performance. The original implementation used numpy MaskedArray, which can be slow. This implementation uses plain ndarray, using np.nan values to represent masked pixels. This is possible due to a change where all non-finite values (i.e, NaN, inf) are automatically masked. This is consistent with the behavior of astropy SigmaClip, which the background classes use heavily.

Additional performance gains also now come when the optional bottleneck package is installed.

Overall, the performance gains can be up to 10-50%, depending on the array size and input parameters.

The background_mesh_ma and background_rms_mesh_ma attributes have been renamed to background_mesh_masked and background_rms_mesh_masked and are now deprecated.

This PR introduces a backwards-incompatible API change in that the background statistics classes (e.g., MeanBackground) now return a ndarray with np.nan values at masked pixels instead of a MaskedArray. Any user-defined bkg_estimator or bkgrms_estimator functions/classes input to Background2D must also have the same behavior.

This PR also adds a masked keyword to the background statistic classes to return a MaskedArray.