Allow for relative, up to 5% differences, while comparing numerics for fieldmap correspondence

[yarikoptic]

The prime target is affines, but I am afraid ATM it would apply to anything numeric and thus potentially be not desired! Original use case triggering this is OpenNeuroDatasets/ds005256#2 .

Also looking back at when we added this, it was in bf9a3e1 and without due description -- we just switched from exact to np.allclose with its default of rtol=1e-05, atol=1e-08, equal_nan=False.

TODOs

• see how to restrict lowered rtol to affines ... wdyt @psadil and @bpinsard ?

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.04%. Comparing base (5b1718b) to head (2567abb).
⚠️ Report is 20 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #842      +/-   ##
==========================================
+ Coverage   82.84%   84.04%   +1.19%     
==========================================
  Files          42       42              
  Lines        4313     4800     +487     
==========================================
+ Hits         3573     4034     +461     
- Misses        740      766      +26     

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[psadil]

I think that'd be great!

As I understand, the kinds of differences that we see in the affines are often not at the scale of numerical precision (which is targeted by the the defaults for np.allclose), but rather some small genuine differences caused in how parameters end up getting set --- genuine but inconsequential. So 5% sounds appropriate to me.