Use zero-copy approach during subkeys message packing

The following discussion from !39 should be addressed:

• @akudiyar started a discussion: (+2 comments)

  I see a potential bottleneck here. Imagine a high load scenario in which we will overload GC with constant byte arrays creation.

  I think that we should somehow use the byte buffer from MessagePack packer and read the values for calculating the hash directly from it (after "flipping" it perhaps). In this case a custom murmur hash implementation can benefit us more than a library.