Evidence handling and submission

[ethanfrey]

If you try to update the client to header H, but a different header was previously submitted, we have a clear case of a byzantine network (> 2/3 voting for different blocks at the same height). In this case, the relayer should be able to repackage it's knowledge into evidence of cheating on the source chain (the one who's header we were submitting, not the chain we were submitting to).

This will be hard to test without a real Byzantine testnet, so this is mainly a placeholder until there is a good test environment for this.