--val --val_best fails when checkpoint_final is missing

Hi,

We had an interrupted training where checkpoint_best.pth was available but there was no checkpoint_final.pth.

I wanted to run the validation with:

nnUNetv2_train DATASET 3d_fullres FOLD --val --val_best

I found #2459 where this command was suggested for this situation. But on the current master (27f878d) I get:

RuntimeError: Cannot run validation because the training is not finished yet 

I also made a small test without data or GPU:

from pathlib import Path
from tempfile import TemporaryDirectory
from nnunetv2.run.run_training import maybe_load_checkpoint

class Trainer:
    def __init__(self, folder):
        self.output_folder = folder
    def load_checkpoint(self, path):
        print(path)

with TemporaryDirectory() as folder:
    Path(folder, "checkpoint_best.pth").touch()
    maybe_load_checkpoint(Trainer(folder), False, True)

It looks like maybe_load_checkpoint checks for checkpoint_final.pth first and raises the error. The code for val_with_best is called only after this, so it never reaches checkpoint_best.pth.

I know that stopping the training early is not the recommended nnU-Net use. I am just not sure if this behavior is expected.

Would it make sense that:

• --val requires checkpoint_final.pth
• --val --val_best requires checkpoint_best.pth

If yes, I can propose a small PR with tests.

Thank you!