Skip call to .tolist() when creating pd.Index

`np.tile` returns an NDArray and there is no need to convert this to a
Python `list` prior to passing it to `pd.MultiIndex`. The interface to
Pandas requires the object to be array-like, and one of the first things
that the constructor does is coerce the list back to an NDArray.

For arrays with large coordinate axes, `to_dataframe()` is extremely
slow due to Pandas needing to iterate through a `list` object rather
than an array.

For an (1000, 500, 20) array -- 10M rows in the cartesian product --
this results in a ~20x speed-up for `xr.Dataset.to_dataframe()` (tested
on x86 and Apple Silicon).

```python
da = xr.DataArray(np.ones((1000, 500, 20)), name="foo")
da.to_dataframe()
```

Closes #10617

Author: y4n9squared

xarray/core/coordinates.py

@@ -177,7 +177,7 @@ def to_index(self, ordered_dims: Sequence[Hashable] | None = None) -> pd.Index:
 
                 # compute the cartesian product
                 code_list += [
-                    np.tile(np.repeat(code, repeat_counts[i]), tile_counts[i]).tolist()
+                    np.tile(np.repeat(code, repeat_counts[i]), tile_counts[i])
                     for code in codes
                 ]
                 level_list += levels