[ipfs/go-bitswap] Improve (CPU) performance when rebroadcasting wants

When we can't find anything for a while, we add all outstanding wants to a "broadcast" wantlist. Unfortunately, we _repeatedly_ do this. Fortunately, we don't re-send "want" messages but repeatedly adding all idle wants to every peer's wantlist incurs a high CPU cost.

We've partially fixed this in ipfs/go-bitswap#133 by doing this less frequently but we can do better.

1. If it's already in the broadcast wantlist, don't bother "sending" the message.
2. Only add _new_ idle wants? Extra state we need to track but may be worth it.