Tpcp

Added

- The global cache helper now support algorithms with multiple action methods by specifying the name of the action
method you want to cache.
(https://github.com/mad-lab-fau/tpcp/pull/118)
- Global disk cache helper should now be able to cache the action methods of algorithm classes defined in the main
script.
(https://github.com/mad-lab-fau/tpcp/pull/118)
- There are new builtin `FloatAggregator` and `MacroFloatAggregator` that should cover many of the use cases that
previously required custom aggregators.
(https://github.com/mad-lab-fau/tpcp/pull/118)
- Scorers now support passing a `final_aggregator`. This is called after all scoring and aggregation happens and allows
to implement complicated "meta" aggregation that depends on the results of all scores of all datapoints.
Note, that we are not sure yet, if this should be used more as an escape hedge and overusing it should be considered
an anti-pattern, or if it is exactly the other way around.
We need to experiment in a couple of real-life applications to figure this out.
(https://github.com/mad-lab-fau/tpcp/pull/120)
- Dataset classes now have a proper `__equals__` implementation.
(https://github.com/mad-lab-fau/tpcp/pull/120)

Changed

- Relative major overhall of how aggregator in scoring functions work. Before, aggregators were classes that were
initialized with the value of a score. Now they are instances of a class that is called with the value of a score.
This change allows it to create "configurable" aggregators that get the configuration at initialization time.
(https://github.com/mad-lab-fau/tpcp/pull/118)
This comes with a couple of breaking changes:
- The most "user-facing" one is that the `NoAgg` aggregator is now called `no_agg` indicating that it is an instance
of a class and not a class itself.
- All custom aggregators need to be rewritten, but you will likely find, that they are much simpler now.
(see the reworked examples for custom aggregators)

Fixed

- Fixed massive performance regression in version 0.34.1 affecting people that had tensorflow or torch installed, but
did not use it in their code.
The reason for that was, that we imported the two modules in the global scope, which caused importing tpcp to be very
slow.
This was particularly noticeable in case of multiprocessing, as the module was imported in every worker process.
We now only import the module, within the clone function and only, if you had imported it before.
(https://github.com/mad-lab-fau/tpcp/pull/118)
- The custom hash function now has a different way of hashing functions and classes defined in local scopes.
This should prevent strange pickling errors from just using "tpcp" normally.
(https://github.com/mad-lab-fau/tpcp/pull/118)

Removed

- `score` functions implemented directly as method on the pipeline class are no longer supported.
Score functions now need to be independent functions that take a pipeline instance as their first argument.
For this reason, it is also no longer supported to pass `None` as argument to `scoring` in any validate or optimize
method.
(https://github.com/mad-lab-fau/tpcp/pull/120)

Fixes names of optional dependency groups. That should resolve install issues when using uv as package manager.

Note: This is a major version bump, because we have quite substantial breaking changes. The 1.0 should not signal that we
are now feature complete. Though the core APIs have been mostly stable for quite some time now.

BREAKING CHANGE

- Instead of the (annoying) `mock_label` and `group_label` arguments, all functions that take a cv-splitter as input,
can now take an instance of the new `DatasetSplitter` class, which elegantly handles grouping and stratification and
also removes the need of forwarding the `mock_label` and `group_label` arguments to the underlying optimizer.
The use of the `mock_label` and `group_label` arguments has been removed without depreciation.
(https://github.com/mad-lab-fau/tpcp/pull/114)
- All classes and methods that "grid-search" or "cross-validate" like output (`GridSearch`, `GridSearchCv`, `cross_validate`, `validate`)
have updated names for all their output attributes.
In most cases the output naming has switched from a single underscore to a double underscore to separate the different
parts of the output name to make it easier to programmatically access the output.
(https://github.com/mad-lab-fau/tpcp/pull/117)

Fixed
- The torch hasher was not working at all. This is hopefully fixed now.
- The tensorflow clone method did not work. Switched to specialized implementation that hopefully works.

Added

- Dataset classes are now generic and allow you to provide the group-label tuple as generic. This allows for better type
checking and IDE support. (https://github.com/mad-lab-fau/tpcp/pull/113)

Changed/Fixed

- The snapshot utilities are much more robust now and rais appropriate errors when the stored dataframes have
unsupported properties. (https://github.com/mad-lab-fau/tpcp/pull/112)

- Only warning about global caching once.