expansion_factor on tokens should applied to hidden dim not to num_patches

[ease-zh]

#11 has notice the problem about the expansion_factor on tokens.
But, in the official implemtention, and also in the papers, the dimension for tokens mlp is actually half of the hidden dim, and is not relevant to the num of patches.

The most intuitive evidence is that, for the same image resolution, i.e. the same num of patches, the S/M/L/B models has different MLP dimension Ds.