optimize regex handling

[kmod]

I think these are good changes, but the perf results are exactly the same as my non-perf-related PR:

       django_template.py             3.6s (2)             3.6s (2)  -0.2%
            pyxl_bench.py             3.5s (2)             3.4s (2)  -1.2%
sqlalchemy_imperative2.py             4.3s (2)             4.4s (2)  +0.4%
                  geomean                 3.8s                 3.8s  -0.4%

It speeds up some simple regex microbenchmarks by 4x though.

[kmod]

It looks like Travis-CI tested the wrong commit, but I ran the tests manually and it seems ok.

Here are the stats on the benchmarks I was looking at:

         django_lexing.py             2.0s (2)             1.8s (2)  -11.6%
      re_split_ubench2.py             2.9s (2)             1.8s (2)  -38.7%
       re_split_ubench.py             3.1s (2)             0.8s (2)  -75.3%

django_lexing is the lexing part of the template parsing; re_split_ubench2 is the regex portion of the lexing, and re_split_ubench is just a tiny regex (something like re.split("banana" * 100, "ana")).