minor tweak to profiled code

[Matthew Fluet]

> Just  to  make  the profiling assembler output a bit more readable, can't all
> the xxx$$End symbols that have no code (only alignment) between them and  the
> next   xxx$$Begin   be  eliminated?   Wouldn't  this  speed  up  mlprof,  and
> compilation under profiling?  In a test program, the .S file went  from  9009
> lines and 310K to 7175 lines and 205K be eliminating the $$End lines.

Looks like we can completely drop the $$End labels.  mlprof isn't actually
using them for anything.  Previously, they were there for a sanity check:
every instruction produced by the native backend should be in exactly one
begin/end pair -- i.e., the begin/end pairs form a disjoint union that
covers the code segment (corresponding to native codegen code).  It's easy
enough to supress them.