This test works by spinning up a thread with an atypically small stack
size, parsing a regex into an Ast and then dropping it. We use a small
stack size such that *if the Ast didn't have a custom Drop impl*, then
its default recursive Drop impl would overflow the stack. (If we don't
use a smaller stack size, then the default on some platforms is usually
quite large and might require a much larger Ast to provoke a failure.)

It turns out that the stack size we were using was quite tiny, and too
tiny for some platforms such as FreeBSD. We therefore increase it a
little bit, but not too much.

We do the same for the corresponding test for the custom Drop impl for
Hir.

Fixes #967

In turns out that in *some* calls to Fsm::reverse, we were passing an
incorrect start offset. Namely, the haystack we pass is sub-sliced at
`&text[start..]`, but in some places, we were passing `text.len()` as
the start offset of the reverse search. But of course, it should be
`text.len() - start`. This was indeed the case in most places, but it
looks like it needed to be corrected in two additional places.

I've also added this test to regex-automata's set of regression tests
and can confirm that it doesn't happen there. (regex-automata is far
more principled about handling offsets like this.)

Fixes #969