PARSE-RTL now inserts PROFILE-COUNT instructions at the end of basic

PARSE-RTL now inserts PROFILE-COUNT instructions at the end of basic
block.  It mostly works, but this is the wrong place for it.  There is
no effect unless compiler:generate-profiling-instructions? is true.

The instuctions should be added just prior to assembly.  This would
make PARSE-RTL a reusable utility (we dont want two or three counts
per basic block!), and it would forestall problems with RTL
optimizations that change basic block boundaries (like common suffic
merging).