[BUG] Some models do not drop unnecessary stages when specifying out_indices

[abcamiletto]

Describe the bug
Some models do not drop the unused stages when specifying out indices

This might be related to this. I had the impression it happens also with other models, but I might be wrong and this may be be specific to tf_efficientnets.

In any case, thanks a lot for this amazing library!

To Reproduce

import timm
import torch

resnet_full = timm.create_model('resnet18', features_only=True)
resnet = timm.create_model('resnet18', features_only=True, out_indices=(0,1,2))

effnet_full = timm.create_model('tf_efficientnet_b2', features_only=True)
effnet = timm.create_model('tf_efficientnet_b2', features_only=True, out_indices=(0,1,2))

get_n_params = lambda model: sum(p.numel() for p in model.parameters() if p.requires_grad)

resnet_params_full = get_n_params(resnet_full)
resnet_params = get_n_params(resnet)

assert resnet_params_full > resnet_params

effnet_params_full = get_n_params(effnet_full)
effnet_params = get_n_params(effnet)

assert effnet_params_full > effnet_params # Fails

Desktop (please complete the following information):

• OS: [e.g. Windows 10, Ubuntu 18.04]
• This repository version [e.g. pip 0.3.1 or commit ref]
• PyTorch version w/ CUDA/cuDNN [e.g. from conda list, 1.7.0 py3.8_cuda11.0.221_cudnn8.0.3_0]