Weight dropout, var weight noise make checkpoint incompatible

[albertz]

Because the parameter name changes.

I think it would be good if you could directly load a checkpoint with mismatching conditions (weight dropout on or off).

The code should stay simple though. The nice thing about the current code is that it is quite straightforward and simple.

[albertz]

Note that we can also handle this on RETURNN side in our CustomCheckpointLoader, and potentially replace <name>_raw by <name> and vice-versa. This is simple, unlikely to be wrong, and would solve the problem here.