Improve Enzyme in forward mode with chunk size

[gdalle]

Core

• Add a pick_chunksize heuristic lifted from ForwardDiff.jl

Extensions

• Enzyme forward: Use this chunk size selector to prepare gradient and jacobian
• Enzyme reverse: Try to do the same for jacobian but I ran into Reverse mode jacobian is wrong for matrix input EnzymeAD/Enzyme.jl#1391

Tests

• Add AutoEnzyme(Enzyme.Forward) to the type-stable backends

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.75%. Comparing base (de94031) to head (8a84314).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #186      +/-   ##
==========================================
+ Coverage   93.60%   93.75%   +0.15%     
==========================================
  Files          66       67       +1     
  Lines        3222     3237      +15     
==========================================
+ Hits         3016     3035      +19     
+ Misses        206      202       -4     

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

[adrhill]

In the long run, would it make sense to add a backend argument for dispatch to pick_chunksize? Or is the heuristic generally valid?

[gdalle]

It can't hurt. But at the moment only ForwardDiff and Enzyme handle chunks, and based on the slack discussion the optimal chunk length is about the same for both (circa SIMD length)