Performance improvements for Similarity family

[jabooth]

Simple improvements to Similarity transform family gives some nice speed ups to Procrustes.

print('{} {}D pointclouds each of {} '
      'points'.format(len(ps), ps[0].n_dims, 
      ps[0].n_points))
88 3D pointclouds each of 60 points

%timeit GeneralizedProcrustesAnalysis(ps)
296 ms -> 119 ms

%timeit AlignmentSimilarity(ps[0], ps[1])
833 µs -> 317 µs

[patricksnape]

These are failing 😒

[jabooth]

Ah this test was not running on my machine due to numerical instabilities between OS X and Linux. Investigating.

[jabooth]

updated the test suite to pass with the new numbers.

To be honest, the test that failed is highly numerically unstable - it's only useful in flagging up the fact that something numerically unstable has changed in the code that the SDM fitter uses, which is exactly what has happened here.

This isn't surprising, as for efficiency reasons this PR moves from doing

b.apply(a.apply(x))

to

a.compose_before(b).apply(x)

which will lead to slight numerical differences that aren't 'good' or 'bad'.

[patricksnape]

+1 fair explanation