Better buffering

[Stebalien]

We should add buffering to all of our message-based APIs so we avoid sending message fragments. This should alleviate libp2p/go-libp2p#235.

[Stebalien]

I've now tried:

• A "double buffer" buffering library (write to a buffer while another thread flushes).
• Manually adding buffering to bitswap to avoid sending messages lengths separate from messages.

No dice. Nothing makes a difference. The thing is, bitswap doesn't send a lot of little messages (to the same peer). We send a lot of little messages to many peers (buffering won't help) and many large messages to a single peer (buffering won't help). Even if we end up sending the message length in a separate packet/secio-message, that extra message is easily amortized by the size of the primary message in most cases.

TL;DR: Dead end.

[Stebalien]

I take this back. Adding double buffering (with a small delay) helps reduce the number of packets quite a bit (testing with wireshark). I believe the issues with my prior testing were:

1. I was testing bitswap, not stuff like the DHT (tons of small packets). Buffering should help background bandwidth.
2. Localhost is not the right environment to test on. Reducing the number of packets sent is significantly more important on the open internet.

[Stebalien]

Closing in favor of #3155.