Implement ADviaFDConfig

[mzgubic]

Closes #241 and #221

[codecov-commenter]

Codecov Report

Merging #243 (1bfef83) into main (dd0e246) will increase coverage by 0.39%.
The diff coverage is 94.11%.

@@            Coverage Diff             @@
##             main     #243      +/-   ##
==========================================
+ Coverage   90.75%   91.15%   +0.39%     
==========================================
  Files          11       12       +1     
  Lines         303      339      +36     
==========================================
+ Hits          275      309      +34     
- Misses         28       30       +2     

Impacted Files	Coverage Δ
src/ChainRulesTestUtils.jl	100.00% <ø> (ø)
src/rand_tangent.jl	96.29% <85.71%> (-3.71%)	⬇️
src/global_checks.jl	92.00% <92.00%> (ø)
src/finite_difference_calls.jl	100.00% <100.00%> (+2.77%)	⬆️
src/rule_config.jl	69.23% <100.00%> (+58.11%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 487890f...1bfef83. Read the comment docs.

[oxinabox]

should this try and use a rrule/frule first then only fallback to using the FiniteDifferences if that returns nothing?

[mzgubic]

should this try and use a rrule/frule first then only fallback to using the FiniteDifferences if that returns nothing?

Do we ever want to avoid using the rule and make sure to use FD for the inner rule? I don't think so.

I think this means we can just combine the ADviaRuleConfig and ADviaFDConfig into a TestConfig, which does what you describe? And it's not a breaking change because it's not exported?

[oxinabox]

I think this means we can just combine the ADviaRuleConfig and ADviaFDConfig into a TestConfig, which does what you describe?

Yeah, I guess so.

And it's not a breaking change because it's not exported?

It is, because people use it directly. Don't they?
But we can @deprecate_binding ADviaRuleConfig TestConfig false

[oxinabox]

merge when happy

[mzgubic]

@oxinabox FYI there was actually a gross bug in the original implementation.

This TestConfig will be used to test rules for higher order functions f(inner, array) where we don't care that much for the rule for inner and want to do finite differencing for it out of convenience.

However, if we don't check whether a rule exists inside f/rrule_via_ad using the TestConfig, we are actually ignoring the outer rule implementation (for f) entirely, and always pass the tests because finite differencing is done on the outer rule 😅

I will wait for you to reply before merging since this was a significant development (even if not a significant change to the code)

[oxinabox]

oh