Add custom gradient

[james77777778]

Related to #18442

This op doesn't support numpy and torch backends.

I have tried to add support for torch backends but failed.
The reason is that the function signature and ctx mechanism is quite different from torch compared to tf/jax.
Ref: https://pytorch.org/tutorials/beginner/examples_autograd/two_layer_net_custom_function.html

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 75.66%. Comparing base (c8700f4) to head (ce14ad5).
Report is 86 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #19279      +/-   ##
==========================================
- Coverage   80.14%   75.66%   -4.48%     
==========================================
  Files         341      365      +24     
  Lines       36163    39778    +3615     
  Branches     7116     7712     +596     
==========================================
+ Hits        28982    30099    +1117     
- Misses       5578     8013    +2435     
- Partials     1603     1666      +63     

Flag	Coverage Δ
keras	75.51% <100.00%> (-4.48%)	⬇️
keras-jax	59.70% <77.77%> (-3.36%)	⬇️
keras-numpy	54.24% <66.66%> (-2.84%)	⬇️
keras-tensorflow	61.21% <77.77%> (-3.44%)	⬇️
keras-torch	60.32% <55.55%> (-3.55%)	⬇️

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.

[fchollet]

Thanks for the PR. I would not have expected this to be feasible for torch indeed.