Improve Jacobian and Hessian preparation

[gdalle]

Versions

• Bump DI to v0.6.5

DI source

• Encode batch size from pick_batchsize with Val(B) instead of B
• Pass Val(B) to the Jacobian and Hessian preparators with a function barrier. Do the same for sparse Jacobians and Hessians.
• Avoid slicing the whole Jacobian/Hessian when the last block is too big, only slice the last block

DI extensions

• Enzyme: set batch size to Val(16)
• ForwardDiff: set batch size to Val(C) if pre-selected in AutoForwardDiff. I tried to do the same for Chunk{C} but it is forbidden to have chunks larger than the input size.

DIT source

• Add a secret option preparation_type_stability to DIT.test_differentiation which runs JET.@test_opt on the preparation operator

DI and DIT tests

• Test preparation_type_stability for trivial dense and sparse backends

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.63%. Comparing base (973676a) to head (09c206d).
Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #535      +/-   ##
==========================================
- Coverage   98.64%   98.63%   -0.02%     
==========================================
  Files         106      106              
  Lines        4572     4606      +34     
==========================================
+ Hits         4510     4543      +33     
- Misses         62       63       +1     

Flag	Coverage Δ
DI	98.76% <100.00%> (-0.03%)	⬇️
DIT	98.34% <100.00%> (+0.01%)	⬆️

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.