[BUG] Wrong model.num_features for mobilenetv4

[MichaelMonashev]

Describe the bug
model.num_features return 960, but correct number is 1280.

To Reproduce

import timm
model = timm.create_model('mobilenetv4_conv_large.e500_r256_in1k', pretrained=True, num_classes=0)
print(model.num_features)
960

Expected behavior

print(model.num_features)
1280

[rwightman]

@MichaelMonashev it is not wrong https://huggingface.co/timm/mobilenetv4_conv_small.e2400_r224_in1k/discussions/3

num_features (last feature map before pooling) == output of forward_features
head_hidden_size (pre-logits features, before classification head ) from output for forward_head(..., pre_logits=True) or num_classes=0

head_hidden_size != num_features for models like mobilenetv3/v4 or some other models that have mlps or other conv/linear layers in the head AFTER pooling