Add graph showing error in function of CSA

[PaulBautin]

Hypothesis is that Deepseg is less sensitive to rescaling on subjects with small CSA (atrophy under-estimation or CSA over-estimation).

[PaulBautin]

Results for T2w images using: config without resampling: config_script_t2.zip

We can observe a propensity for smaller CSA to have a larger error. For this graph, because we know that error increases with 'CSA scaling', I have normalized mean CSA (dividing mean CSA by 'CSA scaling'). This allows direct comparison of mean CSA across scalings. Ideally, without any biases (increase in error) for small CSA, the dotted line should be horizontal.