Fixed #24743, #24745 -- Optimized migration plan handling

The change partly goes back to the old behavior for forwards migrations
which should reduce the amount of memory consumption (#24745). However,
by the way the current state computation is done (there is no
`state_backwards` on a migration class) this change cannot be applied to
backwards migrations. Hence rolling back migrations still requires the
precomputation and storage of the intermediate migration states.

This improvement also implies that Django does not handle mixed
migration plans anymore. Mixed plans consist of a list of migrations
where some are being applied and others are being unapplied.

Thanks Andrew Godwin, Josh Smeaton and Tim Graham for the review as well
as everybody involved on the ticket that kept me looking into the issue.