GPU support for gradient analyzers

[adrhill]

Instead of computing the gradient of the input w.r.t. an indexed model output, this PR computes a vector-jacobian-product w.r.t. to the full model output, selecting the vector v such that it masks the model model outputs.

Besides working well with CUDA.jl, this avoids a second forward-pass through the model and greatly simplifies the code by removing the distinction between gradient_wrt_input and gradient_wrt_batch.

[codecov]

Codecov Report

Patch coverage: 100.00% and project coverage change: +0.24% 🎉

Comparison is base (3aa10ba) 94.92% compared to head (dead62e) 95.16%.
Report is 5 commits behind head on master.

❗ Current head dead62e differs from pull request most recent head 9c3fe0b. Consider uploading reports for the commit 9c3fe0b to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #144      +/-   ##
==========================================
+ Coverage   94.92%   95.16%   +0.24%     
==========================================
  Files          18       19       +1     
  Lines         689      683       -6     
==========================================
- Hits          654      650       -4     
+ Misses         35       33       -2     

Files Changed	Coverage Δ
src/ExplainableAI.jl	100.00% <ø> (ø)
src/lrp/rules.jl	94.16% <ø> (+0.96%)	⬆️
ext/TullioLRPRulesExt.jl	100.00% <100.00%> (ø)
src/gradient.jl	100.00% <100.00%> (ø)
src/lrp/canonize.jl	100.00% <100.00%> (ø)

... and 1 file with indirect coverage changes

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