Allow Activation Layer of MLP In SwinTransformer to be customizable

🚀 The feature

Currently we do

self.mlp = MLP(dim, [int(dim * mlp_ratio), dim], activation_layer=nn.GELU, inplace=None, dropout=dropout)

But according to original implementation it is

https://github.com/microsoft/Swin-Transformer/blob/afeb877fba1139dfbc186276983af2abb02c2196/models/swin_transformer.py#L197

    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 act_layer=nn.GELU, norm_layer=nn.LayerNorm,
                 fused_window_process=False):

Motivation, pitch

I think it might be better to align with the original implementation by creating extra init param activation_layer and allowing the flexibility.

Alternatives

No response

Additional context

No response

The issue identifies the SwinTransformer MLP construction and links the original Swin Transformer implementation; start by locating that construction in the project and comparing its initializer parameters. The change is complete when the activation layer can be supplied through the SwinTransformer interface while preserving the current default behavior, with existing relevant tests still passing.