Fix SingleMemoryStorageSchedule checkpoint clearing during forward replay

[sghelichkhani]

Summary

Fixes #211.

During forward replay, the checkpoint clearing logic for SingleMemoryStorageSchedule uses adjoint_dependencies from the previous timestep to decide whether a checkpoint is safe to discard. But adjoint_dependencies is only fully populated during the reverse pass (after PR #210), so at forward replay time it can be incomplete, causing needed checkpoints to be cleared.

This replaces the adjoint_dependencies check with a block variable identity check:

if var.output.block_variable is var:
    var._checkpoint = None

If var.output.block_variable is var, the Function still holds the correct value in memory, so the checkpoint copy is redundant. If is not var, a later block has taken ownership and will overwrite the Function during replay, so the checkpoint must be kept.

Test

Tested against the MFE on issue #211.