Investigate opportunities for auto-vectorization

[Ostrzyciel]

In #373 we've introduced a few loops for batch initialization of objects, and what do you know, they compile nicely in AVX instructions. This case has rather minimal impact on performance (the code is executed few times), but a similar principle could be applied elsewhere too.

A good case for that would be this: Jelly-RDF/jelly-protobuf#41

But it could be generalized to have the encoder process minibatches of triples (say, 8–16 triples at a time). This would reduce the number of calls in the stack and allow us to write mini-loops of 8–16 operations everywhere.

Note that 16*4B per pointer = 64B, which is often one cache line in L1.

The widest ops on x86-64 are 512-bit wide, which is exactly 64B. Though, these tend to be not-so-efficient, and I think like literally one CPU line actually implements these in a single clock.

[Ostrzyciel]

Minibatches of things could be buffered via fixed-size arrays, where we'd simply do arraycopy between them – we could even have a nice interface for that to wrap it. This arraycopy would be, on an AVX CPU, simply moving a register.

[Ostrzyciel]

This may be more applicable to the encoder than the decoder, as in the decoder we have much less control over the input to the processing loop. Also, optimizations like the one in #396 will conflict with that, so we must be careful to consider if this is really worth it there.