[Bart/Memory] SelfAttention only returns weights if config.output_attentions

[sshleifer]

Previously, SelfAttention would always return attn_weights, and then BartDecoder and BartEncoder would decide whether to return them to the user.
The attn_weights tensor is fairly large, with shape = (bs, num_heads, tgt_len, src_len)
This meant that the memory allocated for attn_weights could not be freed until after the forward pass of BartDecoder.

Now: SelfAttention returns (output, None) if config.output_attentions=False and the memory can be freed

Impact: memory can be freed after SelfAttention returns. -600MB peak GPU consumption for batch_size=6, tgt_len=src_len=1024, num_heads=16

Speed impact: negligible

[thomwolf]

Nice!