Rearrange coding in COPY so that expansible string buffer for data being

read is reused for successive attributes, instead of being deleted and
recreated from scratch for each value read in.  This reduces palloc/pfree
overhead a lot.  COPY IN still seems to be noticeably slower than it was
in 6.5 --- we need to figure out why.  This change takes care of the only
major performance loss I can see in copy.c itself, so the performance
problem is at a lower level somewhere.