fix: upgrade Mooncake compat to v0.5

[gdalle]

• Bump Mooncake compat to v0.5.1
• Take Config as a whole instead of parsing it
• Remove useless _righttype caches
• Add tests with StaticArrays for new friendly tangents mode

Supersedes #960

See also chalk-lab/Mooncake.jl#979

Note that this still uses internal conversion machinery from Mooncake

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.21%. Comparing base (a266e80) to head (6639952).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #961      +/-   ##
==========================================
- Coverage   98.21%   98.21%   -0.01%     
==========================================
  Files         135      135              
  Lines        8002     7995       -7     
==========================================
- Hits         7859     7852       -7     
  Misses        143      143              

Flag	Coverage Δ
DI	98.97% <100.00%> (-0.01%)	⬇️
DIT	96.22% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ 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.

[Technici4n]

This is OK but you should at least carry over the dy_righttype changes, it never made sense to do that and breaks friendly tangents.

[Technici4n]

I guess this is fine if you just want to get this out of the door!

[gdalle]

What do you mean by "never really made sense to do that"?

[gdalle]

Perhaps we should just merge the Mooncake PR and do everything at once

[Technici4n]

What I mean is that _copy_output is just a copy, it doesn't change the object's type, so I don't understand why it is there unless the intention is specifically to make a copy. This PR is fine and should fix Mooncake CI (including docs).

[AstitvaAggarwal]

looks good, thanks!

[gdalle]

Let's just get https://github.com/chalk-lab/Mooncake.jl/pull/980/changes merged and we'll clean up the DI bindings as we bump our compat bound to Mooncake v0.5.1

[gdalle]

So I tracked down the reason behind the existence of dx_righttype and dy_righttype to #513. At first, it was just a way to safeguard against the user providing, say, x::Vector{Float64} and dx::Vector{Float32}, similar to what is done in the Enzyme extension. For lack of friendly tangents infrastructure, the best one could do is make sure that the primal type was equal to the tangent type, and this would work for simple numbers and arrays (which is all I test on).

Initially, this used the handrolled function

copyto!!(dst::Number, src::Number) = convert(typeof(dst), src)
copyto!!(dst, src) = copyto!(dst, src)

But then, in #809, I replaced this function with the one from Mooncake, neglecting the fact that it doesn't accept inputs of different (el)types. Which means the functionality has been essentially useless for a while and we should remove it during the switch to Mooncake 0.5, now that friendly tangents take care of conversion.

[Technici4n]

I see, that makes sense, thanks for investigating. ^^

[gdalle]

Tests currently fail in forward mode with friendly tangents and in-place function because the dy provided by the user is not necessarily the one that gets modified. I think I have to use the same trick as for reverse mode and define a new

f!_with_return(y, x) = begin
    f!(y, x)
    return y
end

[sunxd3]

Thanks for the work!