logit_bias init for non siglip pretrained models

[rsomani95]

Fixes #712

There's one awkward bit about this currently:

import open_clip, torch

m = open_clip.create_model("convnext_base_w", "laion_aesthetic_s13b_b82k", init_logit_bias=100)
m.logit_bias == torch.tensor(0.)   # Maybe this is ok, but this feels awkward?

[rwightman]

@rsomani95 I think it should be

    if 'logit_bias' not in state_dict and model.logit_bias is not None:
        state_dict["logit_bias"] = torch.tensor(0.)

EDIT: actually maybe state_dict["logit_bias"] = torch.zeros_like(state_dict["logit_scale"]) in case we allow loading state dict onto specific device or in diff precision in the future