Make matching of R and native call stacks more robust

[krlmlr]

@atheriel: Would you like to share your experience?

[atheriel]

Yes, here are a few observations from my experience:

• Matching on Rf_eval itself sometimes has issues because that function actually does slightly more than just evaluate R code (e.g. finding the function). As a result, I tried to match only on the combination of Rf_eval + R_execClosure, which is actually where R pushes the new function onto its own stack.

• Some "generic" functions are dispatched in C code (by tryDispatch) and skip the Rf_eval entirely, so I also had to match on those as well.

• Some R "functions" are handled by the interpreter itself and do not appear in the Rprof stack at all, but show up in the C stack. For example c() usage appears with the do_c C symbol, and <- appears with the do_set C symbol.

• I could only get sensible mixed-mode stacks if I deduplicated R-level calls to eval(). I was never able to determine why.