Flexible UNet implementation with backbones

[Nic-Ma]

Is your feature request related to a problem? Please describe.
Followup discussion of #896 .
Currently, we don't have a flexible segmentation pipeline that can support different backbones and pretrained weights.
I think it would be very useful to provide such networks, at least for the UNet implementation.
Actually, we already have some similar prototype in the tutorial:
https://github.com/Project-MONAI/tutorials/blob/main/kaggle/RANZCR/4th_place_solution/models/seg_model.py
I think we can start from integrating this implementation.
@wyli @yiheng-wang-nv What do you think?

[wyli]

To be clear about the requirements, there are multiresolution feature encoders implemented such as

MONAI/monai/networks/nets/efficientnet.py

Line 559 in 78509ae

class EfficientNetBNFeatures(EfficientNet):

This ticket looks for

• a unet-style decoder, for multiresolution upsampling/deconv and skip connections
• segmentation head as the final output block of decoder
• A wrapper for encoder/decoder/head, this could be a generic segmentation wrapper

[yiheng-wang-nv]

Hi @Nic-Ma and @wyli , sure. @binliunls and I will work on this task together.