Redundant training config parameter ffn_hidden in the case of SwiGLU #276

The training config parameter ffn_hidden, see e.g. here, currently needs to be specified (as a multiple of 128) although it is only actually used in the case of GeLU. In contrast, if SwiGLU is used, the effective hidden dimension is automatically derived from n_embd instead, see here.

This is misleading and should be fixed.