Attached is the final (I hope) version of the changes I've been
pushing to the "declarednull" branch in git.  tl;dr with this option
set "typeset foo" leaves foo unset, in contrast to the default
behavior which assigns foo="".

For those who've lost track, this originated from the thread "Bug with
unset variables" in workers/47351 and the first pass at this approach
appeared in workers/47697.  The doc update includes mention of the
problems with searching for (or in) empty strings, from the thread at
workers/47729, and the new test file E03posix.ztst has "xfail" tests
for nine incompatibilities that were recently enumerated in a number
of messages on the list.  I considered making a couple of patches but
one would have been very small and the other would patch a
newly-created file, so it didn't seem worthwhile.

I chose TYPESET_TO_UNSET as the option to control this, to go along
with TYPESET_SILENT and because it works a bit better when negated as
NO_TYPESET_TO_UNSET.  E03posix.ztst runs with this option set to avoid
having to duplicate all the tests; that can be fixed later if
necessary.

I documented the default initialization behavior to contrast it with
the TYPESET_TO_UNSET behavior, rather than try to explain the latter
in isolation.