Avoid allocation for each serialization [Feature]

I love using `rmp_serde`, great work.

I'm interested in very high-performance client-server interactions.

One thing that bothers me about `rmp_serde` is that allocation is forced for each serialization (`to_vec` internally creates its own vector). This lowers performance, if I serialize millions of messages, I could have a pool of buffers amortizing allocation costs and increasing data-cache hits inside the CPU (for the later step when these buffers are read).

Would it be feasible to add a function like `to_existing_vec(val: &T, buffer: &mut [u8])`?

I am not familiar with the ownership mechanics inside the serializer, so I understand if this is turns out to actually be really hard or impossible.