Optional support for nalgebra types

[adamreichold]

This is loosely based on https://github.com/qsib-cbie/nalgebra-numpy (which is based on https://github.com/fusion-engineering/nalgebra-numpy targetting an older version of pyo3) but I tried to achieve a deeper integration into our existing types and optimized the ToPyArray implementation to copy contiguous data using ptr::copy_nonoverlapping.

Closes #263

[davidhewitt]

Agreed that having this as an optional feature makes sense for now. If users wanted to make the ndarray backend optional there can be a longer design discussion to figure out how to make it switchable.

For this PR, do you think the method names should have nalgebra in them? E.g. try_as_nalgebra_matrix() or a new type wrapper .nalgebra().try_as_matrix()?

[adamreichold]

For this PR, do you think the method names should have nalgebra in them? E.g. try_as_nalgebra_matrix() or a new type wrapper .nalgebra().try_as_matrix()?

I think this would be inconsistent with how we handle ndarray (as_array_view instead of the slightly stuttering as_ndarray_array_view) but this is not a particularly strong argument since we do handle both dependencies differently in any case.

I do think that the probability of confusion is small though as ndarray's "array" terminology is nicely distinct from nalgebra's "matrix" terminology.

One major issue I see is discoverability, e.g. searching for nalgebra will not yield those methods. But I think I would prefer to mitigate this using doc aliases:

I will will also extend the top-level documentation to give an overview of the nalgebra support as introduced here.

[kngwyu]

Thank you for this hard work!
I'm new to the nalgebra crate, but I enjoyed learning its unique features while reviewing 🙂
I have a few suggestions:

• Note that the converted NumPy array from nalgebra has always fortran-layout
• Conversion tests for a few more aliases (e.g., Vector3)
• Maybe we might want to ask some original authors (e.g., @de-vri-es) for a quick review if it's OK? I

[adamreichold]

* Note that the converted NumPy array from nalgebra has always fortran-layout

Will do.

* Conversion tests for a few more aliases (e.g., `Vector3`)

Will do.

(I do see room for future improvements here as we currently convert a Vector3 into a 3x1 array with dimensionality two instead of an array with dimensionality one. But I wanted to keep things simple for now.)

* Maybe we might want to ask some original authors (e.g., @de-vri-es) for a quick review if it's OK? I

Certainly. I think pinging them via your comment should be enough, right?

[adamreichold]

Will do.

Done.

[de-vri-es]

Honestly, it's been a while since I've looked at the numpy memory layout. I would trust someone who has dug through the docs more recently more than me ;)

[kngwyu]

I would trust someone who has dug through the docs more recently more than me ;)

Thanks. Fair enough 😅

[mejrs]

Thanks :) just some comments and nits from me.

[kngwyu]

Thank you!