Yet another major improvement for #2799.

Achieve a 2x to 5x boost in isel performance when slicing small arrays by int, slice, list of int, scalar ndarray, or 1-dimensional ndarray.

```python import xarray

da = xarray.DataArray([[1, 2], [3, 4]], dims=['x', 'y']) v = da.variable a = da.variable.values ds = da.to_dataset(name="d")

ds_with_idx = xarray.Dataset({ 'x': [10, 20], 'y': [100, 200], 'd': (('x', 'y'), [[1, 2], [3, 4]]) }) da_with_idx = ds_with_idx.d

before -> after

%timeit a[0] # 121 ns %timeit v[0] # 7 µs %timeit v.isel(x=0) # 10 µs %timeit da[0] # 65 µs -> 15 µs %timeit da.isel(x=0) # 63 µs -> 13 µs %timeit ds.isel(x=0) # 48 µs -> 24 µs %timeit da_with_idx[0] # 209 µs -> 82 µs %timeit da_with_idx.isel(x=0, drop=False) # 135 µs -> 34 µs %timeit da_with_idx.isel(x=0, drop=True) # 101 µs -> 34 µs %timeit ds_with_idx.isel(x=0, drop=False) # 90 µs -> 49 µs %timeit ds_with_idx.isel(x=0, drop=True) # 65 µs -> 49 µs ```

Marked as WIP because this commands running the asv suite to verify there are no regressions for large arrays. (on a separate note, we really need to add the small size cases to asv - as discussed in #3382).

This profoundly alters one of the most important methods in xarray and I must confess it makes me nervous, particularly as I am unsure if the test coverage of DataArray.isel() is as through as that for Dataset.isel().