Subvert GHC's full-laziness optimisation by explicitly lambda-lifting
a local lambda expression to the top level, and exporting it to ensure
it doesn't get inlined.  Without this hack, full-laziness will float
out several subexpressions, which turns out to be a pessimisation in
this case.

This is worth about 20% in hPutStr performance, so we now beat the old
I/O library on 'cat' ;-)

I'd like to MERGE TO STABLE this, but 5.02 is imminent so it might
have to wait until 5.02.1.