Explicit gradients for an atomistic model

[PicoCentauri]

When running a metatensor.atomistic.Model in a simulation engine the model predicts the energy and returns it. Metatensor calls the backward on these energies with respect to positions to get the forces. However some models may want to provide explicit forces where there is no need to call backward by metatensor. This applies to non-conservative force models, for models that may have to break the computational graph or just models that compute forces the old fashioned way.

To implement this I see two ways.

1. The model developer decides: The model adds a TensorBlock as position gradient to the returned TensorMap. The Metatensor engine interface recognizes this and skips the backward
2. The user decides this: We add an option explicit_gradient as part of the ModelOuptut. If this set to true by the user the model knows that it should compute the gradients by itself.

[Luthaf]

The user decides this: We add an option explicit_gradient as part of the ModelOuptut. If this set to true by the user the model knows that it should compute the gradients by itself.

For clarity: this is already how things works: https://docs.metatensor.org/latest/atomistic/reference/models/metadata.html#metatensor.torch.atomistic.ModelOutput.explicit_gradients

ModelOutput.explicit_gradients is a list of string. When used inside ModelCapabilities, this means the model CAN compute gradients explicitly. When used in ModelEvaluationOptions, this means the engine requests the model to compute these. What's mainly missing is a way for the end user to tell the MD engine to request explicit gradients.

---

This applies to non-conservative force models

For this specific case, I would rather have a separate output (since that's how one train such model); and since simulations with non-conservative forces require additional care, I would prefer to make it clear when something is using it.