Garbage collection doesn't distinguish consumers which need checkpoint
files, such as backup, and the ones, who only need WALS, such as
replicas. A disconnected replica will 'hold' all checkpoint files, created
after it got unsynchronised, even though it doesn't need them, which may
lead to disk space shortage. To fix this, we store consumer's type, and
treat consumers differently during garbage collection: now only the old
WALS are stored for replicas, and old checkpoints are stored for backup,
if any. Also changed the tests to check updated garbage collection correctly.

Closes #3444