Add some optimizations to utf8 decoding

[johnynek]

This does four things:

1. adds Chunk.Queue.startsWith so we can be a bit more precise when checking for utf8 byte order mark
2. be miserly in internal methods with allocations and use null instead of Option (avoid allocation of Some on a loop).
3. use a Builder to avoid having to return a tuple and to reverse a list in an internal method.
4. avoid fold on an internal loop and just write out the while loop.

[non]

This looks good. Hopefully you got some great performance improvements from fewer allocations and less memory pressure?

I had a few small suggestions but none of them really affect the correctness of the PR, so take them or leave them.

[non]

I'm pretty sure res is already 0 so i think you just want if (c != counter) res = counter + 1 don't you?

[non]

If bs.size is 1 then what would happen? It seems like minIdx would be set to 0 (0.max(-2)) and idx would also be set to 0 (1-1). In that case, idx = 0 won't break you out of the loop. I think idx = -1 is a safer way to do this (although I would just use return myself).

Oh I see, you're counting on the decrement operator that occurs later. That makes sense, although IMO using -1 or return here might be a bit more reassuring to the reader. (That said, there's no bug so maybe it's not a big deal.)

[non]

It's a small nit but I'd use length instead of size when working with arrays:

scala> val arr = Array(1,2,3)
val arr: Array[Int] = Array(1, 2, 3)

scala> u.reify { arr.size }
val res1: reflect.runtime.universe.Expr[Int] = Expr[Int](Predef.intArrayOps(arr).size)

scala> u.reify { arr.length }
val res2: reflect.runtime.universe.Expr[Int] = Expr[Int](arr.length)

[johnynek]

Benchmark results:

on main:
[info] Benchmark                  (asciiStringSize)   Mode  Cnt       Score       Error  Units
[info] TextBenchmark.asciiDecode                128  thrpt    6  195334.730 ± 18599.880  ops/s
[info] TextBenchmark.asciiDecode               1024  thrpt    6  153746.264 ± 61576.197  ops/s
[info] TextBenchmark.asciiDecode               4096  thrpt    6  118761.218 ±  4695.784  ops/s

on this branch:
[info] Benchmark                  (asciiStringSize)   Mode  Cnt       Score      Error  Units
[info] TextBenchmark.asciiDecode                128  thrpt    6  218236.471 ± 1145.261  ops/s
[info] TextBenchmark.asciiDecode               1024  thrpt    6  192552.031 ± 1689.683  ops/s
[info] TextBenchmark.asciiDecode               4096  thrpt    6  146657.477 ±  589.035  ops/s

So this is 11% to 25% faster depending on the size, for this simple benchmark.