UxDataset.to_xarray() and UxDataArray.to_xarray() should use `_construct_direct`

[Sevans711]

Converting uxarray objects to xarray shouldn't require running all of xarray's validation checks during xr.DataArray.__init__ and xr.Dataset.__init__, because every UxDataArray / UxDataset is already a valid xr.DataArray / xr.Dataset. These checks scale with number of coordinates and variables, and they don't take extremely long so this probably isn't a huge priority, but the overhead could become noticeable in any hot loop handling bookkeeping tasks like this.

For example, in some local tests I swapped UxDataset.to_xarray() to xr.Dataset._construct_direct, and measured speedups (via %%timeit) of:

UxDataset size	main	improved
100 data_vars, 2.8 GB, 150k faces	26 ms ± 169 μs	4.65 μs ± 31.1 ns
265 data_vars, 4.1 GB, 150k faces	118 ms ± 614 μs	5.27 μs ± 14.4 ns

The new UxDataset.to_xarray() implementation was like this:

    def to_xarray(self, grid_format: str = "UGRID") -> xr.Dataset:
        if grid_format == "HEALPix":
                ds = self.rename_dims({"n_face": "cell"})
        else:
            ds = self

        return xr.Dataset._construct_direct(
            dict(ds._variables),
            set(ds._coord_names),
            dims=dict(ds._dims),
            attrs=dict(ds.attrs),
            indexes=dict(ds._indexes),
        )

(Claude suggested this implementation, but I read/understand it, and ran the timing tests myself.)