fix(config): tie spliced set_env markers to their value - #4107

_iter_needs_replacement, which expands a {[section]set_env} splice,
wrote the PEP-508 marker outside the branch that picks the winning value,
so a line that lost the key still installed a marker and suppressed the
value that won. Because the marker is installed lazily during iteration,
the key was present on the first read of the env and gone on every read
after, so one run disagreed with itself.

__init__ and _parse_dict both place the marker write inside the winner
branch; tox-dev#4077 fixed those two and left the splice path alone. The
siblings' `else: markers.pop(...)` is not copied here because it is
unreachable in this parser: once a key is in self._raw every later
spliced line with that key loses, so the branch runs at most once per key.

for more information, see https://pre-commit.ci

Inside one {[section]set_env} block, an earlier conditional line kept
its marker after a later unconditional line or a file| entry redefined
the key, so the inheriting env dropped a variable its parent set. Clear
the marker when the winning line has none, and replace the unit tests
with real runs covering the inherited, redefined, file and still-gated
cases.