Feature/tensor support

[mdbenito]

Description

This PR adds support for tensor data to pydvl.valuation.dataset.Dataset through generics, an Array prototype and a collection of wrapper array functions in pydvl.utils.array.

‼️Beware of the AI slop‼️. This is an attempt at using claude code to implement a complete feature using serena. Despite careful design of tasks and subtasks, and continuously hand-holding the dummy, there were tons of bugs, inconsistencies in the use of types and generics, as well as the array utilities, craploads of duplicate and inane tests which nevertheless left important cases out, as well as several subtle bugs. I hope to have removed much of this, but there is some left, mostly in array.py and associated tests.

Changes

• Dataset now supports instantiation with tensors or numpy arrays. The type is preserved
• Makes all valuation methods agnostic to the array type, except for a few exceptions.
• Fixes some issues with Dataset indexing
• Dataset can take memmapped numpy arrays, or memmap them if mmap=True, reducing memory cost per-node.
• Serialization correctly handles memory maps.
• Updated the MSR notebook, which uses a torch model for the utility to load the data as tensors.
• Introduces a new prototype TorchSupervisedModel, which is implemented e.g. by skorch.NeuralNetClassifier, and used in the MSR notebook (not a new dependency)
• Introduces a new SkorchSupervisedScorer to handle skorch models.

TODO

• Use generics consistently / simplify in array.py

Checklist

• Wrote Unit tests (if necessary)
• Updated Documentation (if necessary)
• Updated Changelog
• If notebooks were added/changed, added boilerplate cells are tagged with "tags": ["hide"] or "tags": ["hide-input"]