Fix or suppress validity test failures in set operations

[awulkiew]

With this PR I intend to fix or suppress test failures in set operations.

There are several cases failing right now. Most of them are failing correctly due to a bug in sym_difference(): #515 and became visible because of this change: #533.

There is a test case which is different, i.e. difference() case mysql_23023665_6:

red: POLYGON((6 7,18 14, -8 1, 0 0, 18 -8, 6 7), (6 0, -4 3, 5 3, 6 0))
green: POLYGON((2 3,-3 5,-10 -1,2 3))
blue: difference(green, red);

This case is fixed by the first commit in this PR which changes the relate() algorithm to use rescaling just like set operations. In other words to use it for areal geometries. This make relate() just slightly more consistent with set operations. It's because in set operations rescaling done WRT bounding box of both geometries, e.g. 2 MultiPolygons but in relate() called in is_valid() rescaling is done WRT pairs of polygons tested for intersecting interiors.

Because of the above after the change in relate() other case starts to fail vailidity test, i.e.: sym_difference() and union() for MultiPolygons ticket_9081:

The problematic fragment is here:

The result of set operation depends on the size of a box used for rescaling. E.g. if union_() of only these 2 polygons is calculated the result is one polygon instead of two:

So I could pass rescaling policy generated for the whole MultiPolygon instead of a pair of Polygons in is_valid(), this might "fix" this issue or not. But even assuming it would fix this particular issue I could easily prepare a different test case which would fail due to this difference in rescaling. So I think for now I'll conditionally disable validity testing for this test case if rescaling is enabled. Ok?

[awulkiew]

Side note. I tried to enable rescaling for all geometries in relate(), also for linestrings and because of this found out that get_rescale_policy() may throw overflow numeric conversion exception here: https://github.com/boostorg/geometry/blob/develop/include/boost/geometry/policies/robustness/get_rescale_policy.hpp#L81. Because diff is very small but greater than machine epsilon and there is no upper limit to rescaled coordinates. This is one of the test cases in question: https://github.com/boostorg/geometry/blob/develop/test/algorithms/equals/equals.cpp#L93
For convenience:

equals(LINESTRING(-3.2333333333333333925452279800083 5.5999999999999978683717927196994,-3.2333333333333333925452279800083 5.5999999999999996447286321199499),
       LINESTRING(-3.2333333333333325043668082798831 5.5999999999999996447286321199499,-3.2333333333333333925452279800083 5.5999999999999996447286321199499))

EDIT: This fails for Linestrings and for Linestrings rescaling is disabled but I bet I could easily create very small Polygon which would cause the same issue.

[awulkiew]

Regarding other difference() cases that are failing, they all fail due to #515 (they all have the same problematic fragment). So I'm going to disable validity test for them as well as @barendgehrels suggested here: 247f657. Or do you have some other suggestion?

[barendgehrels]

I'm OK with this change. Thank you.

[awulkiew]

Thanks. Now lets see if the tests are green again.