Improve performance of snoc case by mutating Builder’s ArrayBuffer

[jamesdbrock]

When snocing to a Builder which has no right node, instead of allocating a new Builder, we could just mutate the current Builder’s ArrayBuffer.

We could also add an effectful preallocate function which would allow us to grow the Builder’s ArrayBuffer in advance of doing a lot of snocing, in the event that we have a good idea of how much snocing we're about to do.

So when we're doing a build which involves a lot of snocs of bytes, then instead of allocating a new ArrayBuffer for each byte as we do now, we could get close to the ideal performance case of just allocating one ArrayBuffer and then mutating it. And this would keep the same API, but with addition of a preallocate function. And we would still get the same performance advantages for doing a subBuilder with our unbalanced binary tree.

[jamesdbrock]

Further, if we assume that the user will use preallocate intelligently, and we could grow the buffer in the case of overflow by some small factor like 1.1 instead of the usual 2.0 doubling factor.

[jamesdbrock]

This optimization would apply to the put* functions, and it would require the creation of a custom MonadWriter instead of using WriterT, so that we could get mutable access to the Monoid.