Quick fix for test_narr_example_variable_without_grid_mapping test (broken by xarray v0.14)

[jthielen]

With xarray v0.14 (in particular, after pydata/xarray#3234), MetPy's usage of xarray.Dataset.apply in parse_cf results in a duplicated coordinate issue, like that observed in 2fe00b6. It seems like it is a very subtle index issue with xarray, however, it can be avoided entirely through using xarray.merge instead. The likely upstream issue remains, but this at least fixes MetPy for now.

[zbruick]

Pending CI. This will fail on the lack of fixing the Gini files here that #1204 does. I'll let @dopplershift handle merging these, as it will require one of them to go in without Travis passing and then a rebase.

[jthielen]

I think I finally figured out what originally went wrong: the unit_array setter on the MetPyDataArrayAccessor does not account for updating indexes. Fortuitously, before pydata/xarray#3234, the DataArray._indexes attribute was not yet modified, so the indexes were only created after the coordinate scaling. However, that seems to have changed with pydata/xarray#3234 when using apply. So, in order to not have apply attempt to combine both unscaled and scaled indexes (which is what ended up happening), we need to somehow "reset" the indexes on each variable to their scaled values.

This means an alternative fix to using merge here would be to add reset_index to the function that is applied in apply. If that fix is preferred, I can modify this PR to do that instead.

[dopplershift]

I like this approach because it reduces the unique code paths at play here. It really goes now: "If no variables requested, do them all through the same path as we would have done."