Distributor snappy.Decode() allocated memory could be optimized with a pool

[pracucci]

Is your feature request related to a problem? Please describe.
Profiling the distributors heap I can see a significant % of memory allocated by call to snappy.Decode() used by util.ParseProtoReader() to parse the remote write request:

Describe the solution you'd like
Given snappy.Decode() picks in input the dst buffer as well, I think we may optimize it reusing the []byte via a pool (we currently pass nil as dst, so the output buffer is always newly allocated).