Detect when scan is being called recursively and
re-use the same scan state.

On my machine, for a recursion-heavy benchmark, this
results in 44x speed up. This does impose a 4% penalty
on the non-recursive case, which can be removed by
heap-allocating the saved state, at 40% performance penalty
on the recursive case. Either way is fine with me.

R=r
CC=golang-dev
https://golang.org/cl/4253049