Re-use read buffer inside gRPC to reduce ingester CPU usage

[bboreham]

According to my profiles, half the CPU in ingesters is taken up with garbage-collection.
70% of garbage is from recvMsg in gRPC.

There is a note on the code already saying to reuse it:
https://github.com/grpc/grpc-go/blob/f320c793495fc90c222831f1708c18119793e0f8/rpc_util.go#L573-L575

I think this is hard in the general case because people can pull tricks like having the deserialised objects point into this buffer, but specifically at this point in Push() / v2Push() we know we don't retain this memory.

I suspect this will be hard to implement, but filing it anyway in case someone wants a challenge.

This will be less valuable if #1227 is implemented, because the buffers will be much smaller.