we should make from-ucontext backtracing work correctly across most implementations

[alk]

SUBJ. See #1524 and 7c73631

I.e. we currently do fixed number of frames skipping for a number of backtracing methods. Most notably libunwind, but also libgcc and pretty much everything else, except generic_fp. The latter knows how to use registers from ucontext to step from right place.

We should consider something like this. Full backtrace is collected, then ucontext's PC (via GetPC) is used to locate right location and then stack would be "cut" up to this place. It should be somewhat reliable (hopefully). It should replace the logic in cpu profiling handler that automagically adds PC from ucontext (and not necessarily correctly).

Good mini-project for someone to contribute IMO.