Use Tcl_ListObjGetElements instead of Tcl_SplitList.  Aside from being
possibly more efficient in its own right, this means we are no longer
responsible for freeing a malloc'd result array, so we can get rid of
a PG_TRY/PG_CATCH block.

Use heap_form_tuple instead of SPI_modifytuple.  We don't need the
extra generality of the latter, since we're always replacing all
columns.  Nor do we need its memory-context-munging, since at this
point we're already out of the SPI environment.

Per comparison of this code to tuple-building code submitted by Jim Nasby.
I've abandoned the thought of merging the two cases into a single routine,
but we may as well make the older code simpler and faster where we can.