Optimize parseNinjaString

The closures in parseNinjaString seem to confuse go's escape
analysis, causing it to allocate a new stateFunc object on the
heap for every character in the string.  Convert the state
functions to normal functions, and pass a pointer to a state
object to provide access to the variables that were in the
closure.

The range iteration on the input string was iterating over
UTF-8 runes, when all of the rune comparisons for state
changes only look at ASCII bytes.  Switch to iterating
over bytes, stringState will pass through multi-byte
characters in strings untouched.

Together this reduces the time spent in parseNinjaString by
>50%, and the overall execution time by >20%.

Change-Id: I4c926b52b883d51e2f12a1673d03fcc2cfe83445