minimizes the changes to `CollectFields` required for incremental delivery (inspired by graphql#3982) -- but retains a single memoized incremental field plan per list item.

…erred (#4320)

In the "old" duplicating version of incremental delivery,
`collectFields()` was responsible for branching, and it therefore
allowed processing a deferred named fragment even if that named fragment
had already been visited as a regular non-deferred named fragment.

One could have argued against that, as the old version had the concept
of inlining, and so we could have just skipped the named fragment and
considered it inlined, but chose not to do that.

Now that we have an algorithm without duplication of fields, revisiting
a named fragment will have no effect, as within the
`buildExecutionPlan()` step, all the duplicated fields will be removed.
So we can just remove the exception and go back to the version
pre-incremental delivery where we always skip a visited named fragment.

Note that we still only consider a named fragment to have been visited
if it was not marked as deferred, because in the case of overlapping
deferred and non-deferred spreads of the same named fragment, we need to
visit it as the non-deferred spread to actually realize that it is
non-deferred.

[As an aside, looks like I made a mistake in
#3994 => we would never have
wanted to ignore the result of `shouldIncludeNode()` => since we are
removing all ignoring of these conditions with respect to defer, this PR
"fixes" that mistake as well.]