When rewriting the do_rm_work injection, do_rm_work_all had been
removed because it seemed obsolete, as do_build now always triggers
do_rm_work.

However, do_build does not get triggered for all recipes and thus
do_rm_work was not called for recipes that got built only
partially. For example, zlib depends indirectly on
zlib-native:do_populate_sysroot.  Because of that dependency,
zlib-native got compiled, but do_rm_work was never called for it.

Re-introducing do_rm_work_all fixes that by making do_build depend on
do_rm_work_all, which then recursively depends on do_rm_work of all
dependencies. This has the unintended side-effect that do_rm_work then
also triggers additional work (like do_populate_lic) that normally
doesn't need to be done for a build. This seems like the lesser evil,
compared to an incomplete cleanup because it mostly enables the
lighter tasks after do_populate_sysroot.

The real solution would be to have two kinds of relationships: a weak
ordering relationship ("if A and B are enabled, A must run before B,
but B can also run without A") and hard dependencies ("B cannot run
unless A has run before").