space profiling

[Matthew Fluet]

> I  just  realized that the space profiling is definitely going to have a real
> problem with overflowing the bins.  If I have a hot loop allocating  lots  of
> junk,  it  isn't  hard  for one location to be responsible for more than 2^32
> bytes of allocation.  This seems like another  argument  for  going  for  the
> procedure call, and having the linked list use 64 bits worth of counter (plus
> 32 bits for the link).  This way you are safe from wrapping.

Yeah, the overflow might be an issue.  This would require additional
changes to mlprof.