Improve speaker model performance.

[barius]

3 smaller issues:

1. When remove_last_nonlinearity: true in config file, there won't be any dropout in embedding layers. This breaks consistency with older models.

2. According to researches, dropout in embedding layers (or the whole network) has no benefit for embedding learning. So we'll remove them.

3. Test Adam + L2.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.