Support checkpoint method

[jihochu]

checkpoint method is suggested in pytorch, and is generally used to reduce runtime memory usage.
It leads to redundant calculations but also reduces memory consumption, and this trade-off can give some options to each application.

To implement the method in nntrainer, several considerations are necessary.
The biggest issue is that nntrainer only supports the pre-calculated memory planning method now. Managed memory area should be planned before the training phase, instant memories could be allocated in layers or calculation methods, but the checkpoint method needs to manage these instant memories more carefully. (It needs to decide when the memory is allocated/freed).
The other issue is related to the policy. It needs to check how we decide which layer needs to be checkpointed. More checkpoint means much memory reduction, also high CPU usage.

[taos-ci]

cibot: Thank you for posting issue #2162. The person in charge will reply soon.

[github-actions]

This issue is stale because it has been open 90 days with no activity. Remove stale label or comment or this will be closed in 3 days.

[github-actions]

This issue was closed because it has been stalled for 3 days with no activity.