Redefine transfroms as set of partially applied functions

[PiotrCzapla]

This is refined version of the previous proposal #312. I've managed to get rid of the TfmType and is_y by converting the transformations to set of applied functions whose parameters can be changed during execution.

We keep the concept of determ, apply_transform from #312. but we maketfms.determ() returning a partially applied function (almost as we return wrapper to make the printout nicer).
The change is backward compatible to large extend.

This let us express the sz_y (the reason foris_y existance) as follows:

def ap_t(t, x, y):
    nx = t(x)
    ny = t(y, sz=sz_y) # here the default value is set to sz_y
    return nx,ny

The coords transformation (one of reasons for TfmType) looks like this:

def ap_t(t, x, y): # this is how we can express coords transformation for y
    nx = t(x)
    ny = transform_coords(t, y, x.shape)
    return nx,ny

The mask transformation (another reason for TfmType) can be expressed as follows:

def ap_t(t, x, y): # this is how we can express coords transformation for y
    nx = t(x)
    ny = t(y, **TfmParams.CLASS)
    return nx,ny

I've also add an inspection mechanism to transforms to make it easier to reason about random transformations, here is how it looks like

>>> tfms[0]
Trasnforms(...)
# with ComposedTransform(tfms=[
  RandomScale.do(_, sz=16, interpolation=3, zoom=?),
  AddPadding.do(_, pad=2, pad_mode=2),
  RandomRotate.do(_, pad_mode=2, interpolation=3, rp=?, rdeg=?),
  RandomLighting.do(_, b=?, c=?),
  RandomDihedral.do(_, rot_times=?, do_flip=?),
  NoCrop.do(_, sz=16, interpolation=3),
  Normalize.do(_, mean=[0.5 0.5 0.5], stddev=[0.5 0.5 0.5]),
  ChannelOrder.do(_)
])

>>> tfms[0].determ()
ComposedTransform(tfms=[
  RandomScale.do(_, sz=16, interpolation=3, zoom=1.0),
  AddPadding.do(_, pad=2, pad_mode=2),
  RandomRotate.do(_, pad_mode=2, interpolation=3, rp=False, rdeg=1.3287164035703682),
  RandomLighting.do(_, b=-0.02255752415182746, c=0.008637519508592914),
  RandomDihedral.do(_, rot_times=2, do_flip=True),
  NoCrop.do(_, sz=16, interpolation=3),
  Normalize.do(_, mean=[0.5 0.5 0.5], stddev=[0.5 0.5 0.5]),
  ChannelOrder.do(_)
])

[jph00]

There's a new transform pipeline now available in fastai_v1