Any un-handled AWS errors would crash with a shown exception that is
difficult to read. This handler just reformats it to a more readable
error-log before still exiting failure.

We implemented this in a downstream user of the library, but it makes
sense to be in use all the time and provided by us.

It can be useful to deploy a single change set that deletes a stack
using some name, then deploys a new one that uses that name. This
doesn't work when we handle deletes after deploys. Reversing it handles
that use-case better, without any downsides.