Optimised key loading in Blake2b

Optimised key loading in Blake2b

The idea is to avoid the slow loading code in the internal
blake2b_update() function, and avoid the overhead of calling
crypto_blake2b_update().

It's a micro-optimisation that in principle shouldn't matter that much,
but it might help a bit if we repeatedly hash small messages with a key,
as can happen in authenticated key exchanges like Monokex.