create new crate for core-ops

[jandremarais]

Separate the core tensor operations from the core crate to a core-ops crate.

[jandremarais]

@edgarriba do you want the operations currently defined in kornia-core/src/tensor.rs like add/mul/powi/etc to live in kornia-core-ops/src/ops.rs?
If so, how should we keep the ability to call for example tensor.powi(2) since we can't define the impl for a foreign type? Do we need a wrapper struct?

[edgarriba]

possibly wiht a TensorOps trait but let's explore this later. Put a first an initial crate with something simple and we keep iterating

[jandremarais]

Ok. Should I start with something simple that is not yet implemented or something that is already implement for Tensor?

[edgarriba]

maybe the norm operator needed by cosine distance

[jandremarais]

I started with the norm operator but then realised it needed the sum reduce operator. So I did its implementation first.

[jandremarais]

Marked this as ready for review so that the new core-ops crate can be merged without being blocked by all the ops that need to be implemented. I you would rather want this to be a bigger PR with more ops added first, I can change the status to draft and work on it further.

[edgarriba]

@jandremarais any progress here ? we can land step by step and keep improving

[jandremarais]

I will work on this tonight (UTC +2). I will remove the dynamic shape stuff, add docs to sum_elements and put the DimOutOfBounds error in TensorError.

[edgarriba]

@jandremarais sound good -- consider this too apache/arrow-rs#6492 (comment). I might be working on a PR for it during the week. Arrow seems pretty unflexible, so maybe we should go for our full implementation

[jandremarais]

@edgarriba I addressed all of your comments in the latest commits. Open to further feedback if I interpreted anything incorrectly or if there are any new suggestions.

[edgarriba]

LGTM. Make a rebase and all good !