fix: revert np.where to xarray.where when adding vars/ constraints

[lkstrp]

#555 introduced masking and used np.where instead of xarray.where since is is "38x faster than xarray where".

But xarray is slower for a reason, since it aligns operands by dimension name, not just position. np.where requires shapes to match positionally, which is not guaranteed in add_variables or add_constraints. This PR just reverts the changes, we should add tests and release a patch tomorrow.

See https://github.com/PyPSA/linopy/pull/555/changes#diff-1a9bf239769f2af33c90e00a21f7f08e2f40d626ce06ee470bb2daa65ac06a04R788-R789

[FBumann]

@lkstrp Sorry if this causes troubles.