Cb

[sebffischer]

This PR:

• Changes the CI to run once a week

• Changes the callbacks (ctx is no longer a argument of each stage) and improves the implementation.

• The cloneability property of a custom callback created by the user is set depending on whether a deep_clone() method was specified. (cloning an untrained learner with a custom callback is still possible)

• Renames:

  • CallbackTorch -> CallbackSet and accordingly all related functions.

  the reason for the renaming was that previously CallbackTorch (the actual callback) and TorchCallback (the descriptor / wrapper) were easy to confuse.

• Added tests the torch optimizer (which were previously misssing)

• task_types no longer has a default when constructing a TorchLoss (the previous default was to use all existing task types in case nothing was specified).

• The autotest for the parameter set was improved and test for it as well

• Unneeded test helpers were removed

• Added the phash methods for the PipeOps

Some TODOs that are left:

• improve the autotest for the callback
• Move from run_paramtest() to autotest_paramset() to also check that the defaults are specified correctly.

NOTE

• The CI for this PR ill only be valid once Fixes #109 is merged
• adding the partial arguments setting throws some warnings due to torch, I already created a PR better handling of partial argument matching mlverse/torch#1075