The point of this commit is to make the memory allocation statistic
from :set +s in GHCi a lot more accurate.  Currently it uses the total
allocation figure calculated by the RTS, which is only updated during
GC, so can be wrong by an arbitrary amount.  The fix is to the the
per-thread allocation counter that was introduced for allocation
limits.

This required changes to the GHC API, because we now have to return
the allocation value from each evaluation.  Rather than just change
the API, I introduced a new API and deprecated the old one.  The new
one is simpler and more extensible, so hopefully we won't need to make
this transition in the future.  See GHC.hs for details.