Added EfficientNetXL. Changed layer naming convention.

This release includes EfficientNetXL model variants.

Due to huge number of layers in this variant:

• The naming convention had to be changed as we exhausted the entire alphabet.
• The Python's default recursion limit has to be increased upon model creation.

Also more weights have been released: imagenet-21k and imagenet-21k-ft1k pretrained models. The names mimic those from original repository.

Initial release.

Initial weights release for EfficientNetV2 models.