Port upstream fixes to skimage.transform estimate methods

[chrisroat]

Porting the following changes:

• ProjectiveTransform has a consistent failure state when estimate is ill-conditioned
• SimilarityTransform.estimate returns False if ill-conditioned (currently always returns True)
• PiecewiseSimilarityTransform.estimate returns False if any piece is illconditioned (currently always returns True)

Upstream refs:
scikit-image/scikit-image#6207
scikit-image/scikit-image#6211

Closes #199

[GPUtester]

Can one of the admins verify this patch?

[quasiben]

add to allowlist

[quasiben]

Currently failing a small style check:

>>>> FAILED: flake8 style check; begin output
./src/cucim/skimage/transform/tests/test_geometric.py:612:1: E302 expected 2 blank lines, found 1

[quasiben]

Also, Thanks @chrisroat !

[chrisroat]

Currently failing a small style check:

>>>> FAILED: flake8 style check; begin output
./src/cucim/skimage/transform/tests/test_geometric.py:612:1: E302 expected 2 blank lines, found 1

Whoops! Fixed.

[jakirkham]

@grlee77 could you please take a look? 🙂

[grlee77]

Sorry, I did not see this before opening #208 just now! The difference between the two is that gh-208 also changes PiecewiseAffineTransform to store the affine parameters on the GPU when the input arrays are on the GPU. The current behavior of always using the CPU is inconsistent with the other classes. See also comments in #208 (comment)

[grlee77]

I think we can close this one in favor of #208, but to credit @chrisroat, I will modify the commits with a "co-authored-by" statement there (Chris fixed the issues upstream in scikit-image).