[FEA] Refactor isclose tests.

[bdice]

Is your feature request related to a problem? Please describe.
In PR #10203, we discovered there is a pytest parametrization for isclose that is 6x6x6x6 = 1296 tests. #10203 (comment)

We can minimize the number of test cases dramatically while improving our test coverage.

Describe the solution you'd like
We want to cover all of the following cases for "low precision" (1e-5 or more, for unit scale data) and "high precision" (1e-8 or less, for unit-scale data):

• values are within rtol but not atol (tiny atol, but the scale of the values is large enough to catch it with rtol)
• values are within atol but not rtol (tiny rtol, scale of values is close to zero)
• values are outside of rtol and atol (easy case)
• values are inside of rtol and atol (easy case)
• different dtypes cast as expected (which requires the "high precision" case above for value scales on the order of 1)

We can parametrize these in a few (4-6) sets of parameters that provide both data and tolerances, instead of the full matrix of all data and tolerance combinations.

Additional context

• isclose is not commutative between a and b. The formula for isclose(a, b, rtol, atol) is absolute(a - b) <= (atol + rtol * absolute(b)).
• Real-world use cases of isclose often deal with comparing "small" differences, often at a precision greater than can be represented in a float32. That is, comparisons of 1.00000001 and 1.
  • e.g. np.float32(1.00000001) == 1 (indistinguishable float32 representations) but not np.isclose(1.00000001, 1, rtol=1e-10, atol=1e-10) (these values may be considered non-equal and not-close at some higher precision).
  • We can "defensively" choose our test data such that it would fail if the data were cast as float32 instead of float64.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[vyasr]

I'm going to close this in favor of #13593