autopilot stats opens extra yamux stream from server to itself

[tgross]

While debugging the issue described in hashicorp/yamux#143 I discovered that a server will open yamux connections to itself in order to collect autopilot stats. The stack trace when we check out the stream from the pool is as follows:

github.com/hashicorp/nomad/helper/pool.(*ConnPool).RPC(0xc000e02500, {0x369e892?, 0x0?}, {0x3d354c0?, 0xc0025d3860?}, {0x36d8eb2, 0x10}, {0x3466960, 0xc0022ce8c0}, {0x2e06bc0, ...})
	github.com/hashicorp/nomad/helper/pool/pool.go:502 +0x16f
github.com/hashicorp/nomad/nomad.(*StatsFetcher).fetch(0xc000984d80, 0xc001a6efc0, 0xc000779490)
	github.com/hashicorp/nomad/nomad/stats_fetcher.go:68 +0x26b
created by github.com/hashicorp/nomad/nomad.(*StatsFetcher).Fetch in goroutine 181
	github.com/hashicorp/nomad/nomad/stats_fetcher.go:97 +0x39d

This is fairly minor but it's unnecessary overhead by opening the stream every 2 seconds, potentially using one of the 64 idle streams in the pool that could be used for real work, requiring an extra TCP connection to self, and making noise when debugging yamux issues. Ideally we'd use the in-mem codec when making this call to the agent it's originating from.