Fine-tune quantization table coefficeints from DCT data

[kornelski]

Scaling of quantization tables assumes that smaller divisors are always better, but that's not always true due to rounding errors.

An especially obvious case is when DC is quantized in a way that prevents it from hitting maximum value (i.e. 128 isn't evenly divisible by the DC coefficient):

quality = 9

quality = 6

e.g. imagine a source image posterized to 16 levels. For such image a DC coefficient = 16 is ideal, and 8 < DC < 16 are all worse, even though they're used at higher qualities.

I think quantization tables can be made a little better:

1. Gather unquantized DCT coefficients for the entire image
2. For each quantization table coefficient check whether higher quantization divisor gives smaller quantization error

[JosePineiro]

Is the same optimization that https://github.com/mozilla/mozjpeg/issues/81 And similar to https://github.com/mozilla/mozjpeg/issues/182

[jrmsmith]

In my opinion, this is a waste of time. A time that would be better spent implementing #182 and #223 -- which provide the best quantization table and the best DCT coefficients.