Optimize keccak

[chfast]

I reimplemented Keccak main loop. This is only used in ethash cache generation. The time for epoch 200 drops from 1.8 s to 1.3 s.

Do not merge it yet.

[fjl]

It looks like a lot of the overhead is due to Sum copying the state and allocating a temporary buffer for the result. We can avoid the overhead using this function:

func makeHasher(h hash.Hash) hasher {
	// If the hash supports Read to get the sum (like sha3.state), use it.
	// This avoids the overhead of Sum.
	type readerHash interface {
		hash.Hash
		Read([]byte) (int, error)
	}
	if rh, ok := h.(readerHash); ok {
		outputLen := rh.Size()
		return func(dest []byte, data []byte) {
			rh.Reset()
			rh.Write(data)
			rh.Read(dest[:outputLen])
		}
	}
	// Otherwise use Sum.
	return func(dest []byte, data []byte) {
		h.Reset()
		h.Write(data)
		h.Sum(dest[:0])
	}
}

Benchmark difference with this change:

name                      old time/op  new time/op  delta
CacheGeneration-8          764ms ± 1%   579ms ± 1%  -24.22%  (p=0.000 n=20+17)
SmallDatasetGeneration-8  75.2ms ±12%  60.6ms ±10%  -19.37%  (p=0.000 n=20+20)
HashimotoLight-8          1.58ms ±11%  1.55ms ± 8%     ~     (p=0.322 n=20+19)
HashimotoFullSmall-8      4.90µs ± 1%  4.88µs ± 1%   -0.31%  (p=0.013 n=19+18)