Unexpected filetype makes inference fail

[hermancollin]

We just received an email from a collaborator who tried the tool. She tried to apply on TIFF images a model trained on PNG images. nnunet doesn't even complain, it just ignores every file because it is looking for PNGs. Along with #5, this is the kind of errors we really need to catch.

In the dataset.json file contained in the packaged models, the field file_ending specifies which extension nnunet is looking for. When we are copying input images in the temporary directory, we could easily check for the extension and apply conversions when necessary.