Add STARPU_FXT_EVENTS to select at runtime events to record

Again, starting with a code review would be nice.

I was wondering if generating at runtime the mask given to fut_keychange() would prevent some compiler optimizations and thus add an extra overhead, but I didn't observe it and, moreover, since fut_keychange() is defined in an external library, I guess the compiler cannot perform much optimizations here.

Draft MR for now, because I noticed a strange overhead even when traces are disabled with distributed executions, so I would like to be sure it doesn't come from this MR first.