have the llvm jit start emitting some capi-style calls

[kmod]

Specifically, for any attribute accesses inside a try-catch block. Even with the new rewrites added here, there's still some perf penalty to using CAPI exceptions, so as we add more of these we should add some better heuristics for determining when to use them. One idea is to only do this for try-catch blocks that catch AttributeErrors; at some point we'll need to make it profiling-based, since quite a few exceptions get thrown through intermediate frames without try-catch blocks. (But then again, the cost of a c++ exception is roughly proportional to the number of frames it crosses so we might not have to do that to start seeing benefit in those cases.)

I'm getting a lot of variability when testing this change, so I tried benchmarking using the pgo build and it seems like the results are more stable:

      django_template3.py             3.2s (2)             3.2s (2)  +0.0%
            pyxl_bench.py             2.8s (2)             2.7s (2)  -2.6%
sqlalchemy_imperative2.py             3.1s (2)             3.1s (2)  -0.8%
                  geomean                 3.0s                 3.0s  -1.1%

or maybe they are just more favorable :)