Small speedup in samping from ortho and special_ortho groups

[JesseLivezey]

Makes 3 changes:

• Uses dot for norm^2,
• Reuse norm calculation, and
• do multiplications by scalars in a different place.

This should be identical up to floating point associativity differences such as between dot(x, x) and (x*x).sum(). np.allclose is satisfied for everything I've checked although np.array_equal is not.

I timed this on my desktop using only 1 core for dim ranging from 2 to 200. Current code is typically 5-15% slower and this tends to shrink for larger dims.

Notebook that generated the tests is here
https://github.com/JesseLivezey/scipy-ortho-speedup/blob/master/faster_ortho_groups.ipynb

[JesseLivezey]

Is it worth adding a benchmark for these?

[rgommers]

It would be nice to have a benchmark, because it looks like there are none for ortho or special_ortho yet.

[JesseLivezey]

@rgommers added a benchmark!

[rgommers]

Benchmark looks good, and verified that code is equivalent. Merged, thanks @JesseLivezey