Clean up num_key_value_heads accessor

[masahi]

I found myself repeatedly copy-pasting the cryptic code

num_key_value_heads = (
    config.num_key_value_heads is None
    and config.num_attention_heads
    or config.num_key_value_heads
) 

to get the kv head size correctly. Also hit a bug where config.num_key_value_heads is accessed directly but found to be None (vicuna).

I added a helper function to LlamaConfig to avoid these troubles.

@junrushao @LeshengJin

[junrushao]

Thanks! In the new nn.Module API, this could be simplified in __post_init__, but we haven't got enough resources migrating to that one yet