fix integer overflow when parsing Perl-extended named backrefs

[cmazakas]

This issue comes to us as: https://issues.oss-fuzz.com/issues/42512790

Note that Perl seems to reject the regex better than us, flagging \k-- as invalid in the first place.

We seem to treat \g and \k as literally the same, while Perl doesn't. Perl rejects \g-- as unterminated token.

Our code seems to parse \k- appropriately, with the assumption that it's a valid synonym for \g, but then we run into overflow issues when parsing the number generated by the fuzzer which is LONG_MIN.

I'm not going to change the behavior of \k here but instead I simply just tried my hand at squashing the overflows.

[jzmaddock]

I don't think this is correct: i < 0 implies an error occurred when parsing the number (negative indexes are always invalid right?), plus mark_count comes from an unsigned value, so is always > 0. Also if you debug the first test case in the new tests, then i ends up being set to 2, and only fails later "by chance" because there are no subsequent marked sub-expressions.

So I would be inclined to remove this line, and instead put the subsequent overflow checks inside a if(i >=0) block. Actually, I'm not sure what should happen if i = 0 either, but it's not handled at all at present, so I'm assuming that's definitely wrong ;)