[FEATURE] Slimmable model interface

[sdatkinson]

Ship an abstraction for slimmable models.

The main public method that is defined is:

void SetSlimmableSize(const double val)

val is assumed to be between 0.0 (minimum size) and 1.0 (maximum size).

The method is not real-time safe. When it is called, the object (whose class also presumably inherits from nam::DSP) should take any necessary steps to prepare itself to .Process() in a new way.

For example, one could imagine a "container" model inside which there are several pre-loaded NAMs (A1-standard, lite, feather, nano?). Its Process function might look like a pass-through to a "hot" model:

void ContainerModel::Process(NAM_SAMPLE** input, NAM_SAMPLE** output, const int num_frames)
{
  mActive->Process(input, output, num_frames);
}

and similar for other public methods. When the SetSlimmableSize method is called, an internal pointer is set to the appropriate model e.g.:

void ContainerModel::SetSlimmableSize(const double val)
{
  if (val < 0.25) {
    mNano->Reset(GetExpectedSampleRate(), mMaxBufferSize);
    mActive = mNano;
  }
  else if (val < 0.5) {
    mFeather->Reset(GetExpectedSampleRate(), mMaxBufferSize);
    mActive = mFeather;
  }
  else if (val < <0.75) {
    mLite->Reset(GetExpectedSampleRate(), mMaxBufferSize);
    mActive = mLite;
  }
  else {
    mStandard->Reset(GetExpectedSampleRate(), mMaxBufferSize);
    mActive = mStandard;
  }
}

This is different from the interface defined in SlimmableNamDsp in that I'm opting for a "continuous" parameter in order to make the presentation more unified for models of different sizes (the paper used a model with max size of 8 channels for demonstration purposes) as well as to make way for other interesting implementations of slimming instead of channel slicing like I did in the paper.

[jfsantos]

I don't seem to be able to assign this to myself but I'm working on it