New model deepgrow_v2: with multi-head inference

Is your feature request related to a problem? Please describe.

Need to be able to train a unet like model (segresnet,etc) as it exist today Plus in addition add a second head to produce 2 outputs. forground and background which would work as deepgrow mode

This network would be trained in 2 modes:

• Main network mode. train segresnet as is with normal 10 labels head
• Deep grow mode. we need to freeze all or segresnet layer. only allow to train the new second head taking from the encoding / decoding and have a new FC layer with 2 outputs

This feature request is related to the following bugs:

• Finetune new model by freezing all layers except FC
• Deepedit has low resolution and non responsive to new clicks

Describe the solution you'd like
Provide new network architecture with 2 heads. This would allow us to train the same core encoding and decoding layer or the network

Allow for 2 modes of training:

• train all label
• train deepgrow
  Allow 2 modes to run infer:
• Run infer as normal segmentation
• Run deep grow

Describe alternatives you've considered

checkptPath = "/rootpath/train_01/train_model.pt"
checkpoint = torch.load(checkptPath)
model_state_dict = checkpoint.get(self.model_state_dict, checkpoint)
self.network.load_state_dict(model_state_dict, strict=False)

Start by reviewing related issues #1298 and #1299, then trace the existing SegResNet and DeepGrow training and inference entry points referenced by this request. The requested result must support shared encoding and decoding with separate segmentation and DeepGrow heads, selectable training modes, and selectable inference modes.