The old code was a depth first graph traversal that could, under the
right conditions, end up re-exploring the same subgraphs multiple
times, once for each way to arrive at that subgraph at a given depth.

The new code uses a breadth first search to make sure that it only
visits each reachable embedded struct once.

Also add fast path for the trivial case.

benchmark                old ns/op    new ns/op    delta
BenchmarkFieldByName1         1321          187  -85.84%
BenchmarkFieldByName2         6118         5186  -15.23%
BenchmarkFieldByName3      8218553        42112  -99.49%

R=gri, r
CC=golang-dev
https://golang.org/cl/6458090