Conformer: Missing dropout after the self-attention

[albertz]

I noticed that we do not have dropout after the self-attention:

    # MHSA
    x_mhsa_ln = self.self_att_layer_norm(x_ffn1_out)
    x_mhsa = self.self_att(x_mhsa_ln, axis=spatial_dim)
    x_mhsa_out = x_mhsa + x_ffn1_out

This is different to the standard Transformer.
This is also different to the paper.

Originally posted by @albertz in #233 (comment)

[albertz]

This is clearly wrong. I wonder how we should deal with such cases. I have lots of experiments now with the old code, so changing the behavior is very inconvenient for me, to properly track the difference. On the other side, we said that changing the behavior is still ok now, and better now than later esp in a case like this here.

Introducing a special option for this might simplify it a bit but is very ugly.

[albertz]

Ok I temporarily introduced the flag use_dropout_after_self_att. It's really only temporarily. It's default is already the correct variant. It's also not really meant to be for normal use, so not a normal argument but a global class attrib instead. My current experiments set this but I try to make all new experiments without this, and then later we can just remove the flag.