Consider switching from `array.compute()` to `dask.compute(array)`

[trexfeathers]

The conversation on dask/dask#11921, including how relaxed Dask are about it, suggests that they expect 'correct' Dask usages to use the dask.compute() function, rather than the .compute() method of Dask objects. It makes me wonder whether the method approach makes us vulnerable to more breaking changes in future (similar to #6417).

I tried making the switch this afternoon, but it produced a lot of test failures.

• Small value changes, with @ukmo-ccbunney suggests might mean a different order of operations
• Integer results contained in length-1 tuples
• Other failures I didn't investigate

This is not affecting us right now, but I wanted to 'plant a flag' in case we get repeats of #6417 happen in future.