use doubles for filter values used with fft in refpix

[braingram]

The fft filtering (fft_interp_norm) used during the the fill_bad_regions function in the refpix step suffers from loss of precision leading to numerical differences across systems. The aa filter coefficients are defined as float32 (temp_a1 is the start of a number of variables used to construct aa):

jwst/jwst/refpix/irs2_subtract_reference.py

Line 1085 in f866557

temp_a1 = (np.cos(np.arange(elen, dtype=np.float32) *

but are used with a double precision fft result:

jwst/jwst/refpix/irs2_subtract_reference.py

Lines 989 to 992 in f866557

	for it in range(n_iter_norm):
	pp = np.fft.fft(p)
	pp *= aa
	p[:] = np.fft.ifft(pp).real

For each loop iteration aa is cast to double and as the added 32 bits are not defined by the algorithm used for the other bits in aa each iteration accumulates error in the output.

This PR switches aa to double precision to avoid the loss of precision. It is expected that this will change the regression test output.

To take one example, test_nirspec_irs2_detector1[linearity] currently passes on jenkins but fails on github actions:
https://github.com/spacetelescope/RegressionTests/actions/runs/9261424291/job/25476839480#step:25:297
although this tests the output file for the linearity step the differences stem from the refpix step (run immediately before linearity for this test). Although the failure is only a single pixel the difference is above the tolerance:

E          Relative tolerance: 1e-05, Absolute tolerance: 1e-07
E         
E         Extension HDU 1 (SCI, 1):
E         
E            Data contains differences:
E              Data differs at [634, 1532, 4, 1]:
E                 a> 0.088768005
E                 b> 0.08876419
E              1 different pixels found (0.00% different).

Furthermore this failure matches that seen when running the same test locally with a mac m1 or on a system with RHEL 9 (jenkins is running RHEL 7).

With this PR the above test fails in the same way on jenkins, github actions, RHEL9 and on a mac m1.

E            Data contains differences:
E              Data differs at [44, 87, 3, 1]:
E                 a> -0.42095566
E                 b> -0.42092705
E              Data differs at [634, 1532, 4, 1]:
E                 a> 0.088768005
E                 b> 0.08876419
E              Data differs at [863, 1294, 5, 1]:
E                 a> 1.7601109
E                  ?       ^
E                 b> 1.7601409
E                  ?       ^
E              Data differs at [5, 1649, 5, 1]:
E                 a> 0.46575546
E                 b> 0.4657402
E              4 different pixels found (0.00% different).

See https://github.com/spacetelescope/RegressionTests/actions/runs/9273993994/job/25515454758 for the regression test run on github actions (with this PR) and https://plwishmaster.stsci.edu:8081/blue/organizations/jenkins/RT%2FJWST-Developers-Pull-Requests/detail/JWST-Developers-Pull-Requests/1480/pipeline/200 for the jenkins run (note that several nan vs 0 failures due to #8488 are unrelated).

This addresses one of several RHEL9 issues (which appear the largely overlap with github actions vs jenkins differences) revealed during testing for JP-3312.

Checklist for PR authors (skip items if you don't have permissions or they are not applicable)

• added entry in CHANGES.rst within the relevant release section
• updated or added relevant tests
• updated relevant documentation
• added relevant milestone
• added relevant label(s)
• ran regression tests, post a link to the Jenkins job below.
  How to run regression tests on a PR
• All comments are resolved
• Make sure the JIRA ticket is resolved properly

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 57.95%. Comparing base (4179c09) to head (83b41eb).
Report is 2 commits behind head on master.

❗ Current head 83b41eb differs from pull request most recent head 67c779b

Please upload reports for the commit 67c779b to get more accurate results.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8512      +/-   ##
==========================================
- Coverage   57.97%   57.95%   -0.03%     
==========================================
  Files         387      387              
  Lines       38830    38810      -20     
==========================================
- Hits        22513    22493      -20     
  Misses      16317    16317              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[hbushouse]

Looks good to me.