Improving visual weighting code and implementation.

[aous72]

This changes visual weighting in a few ways:

• Replaced std::vector with open_htj2k::visual_weights, which uses std::array<T, 96>; we did this because std::vector needs new/delete, which is unnecessary.
• Replaced double with float because these weights are inaccurate and therefore the precision of double is not needed.
• Historically, when the end user does not specify a quantization step size for irreversible wavelet transform, we set the step size to 2^-bit_depth. In this commit, in lines 1324-1325, we set the quantization step size to 2^-min(16, bit_depth), thus preventing a quantization step size smaller than 2^-bit_depth when the end user does not specify it.

@palemieux Please have a look and let me know if you have any concerns.

[palemieux]

FWIW I am not a big fan of this kind of shim that is used only for initialization purposes. I would instead use std::array[] directly.

[aous72]

Thank you for your comment.
You need to to keep a count of how many entries were used; if not they you rely on the producer and consumer using the same number of entries -- that is the only reason.

[palemieux]

Ok. Since it is not a generic array, I would

• use more semantically accurate terms, e.g., count instead of size
• remove clear() since it is never used and I cannot imagine a situation where you would want to c;lear visual weights.
• replace set_weights() with a constructor
• remove push_back()

[palemieux]

... actually, why not make the visual weight generation function static members of the visual_weights class so that we do not need to expose any accessor methods.

[aous72]

@palemieux I followed you suggestion. Have a look and let me know if you have any concerns, improvements. Thank you.