Add int tensor cast - WIP

[A2va]

Pull Request Template

Checklist

• Confirmed that cargo run-checks command has been executed.
• Made sure the book is up to date with changes in this PR.

Changes

As my project is blocked by both #3262 and a TensorFlow-generated ONNX issue, I decided to try implementing this feature.

My implementation is based on the existing float_cast function, so things might be different for integer ones. It is pretty much a work in progress, so there might be errors, and I look forward to receiving any feedback.

Regarding the current state of the implementation, I have added int_cast to all backends, but I am struggling to implement it for the ndarray backend.
I have added the implementation of IntNdArrayElement (crates/burn-ndarray/src/element.rs), but since all uxx are unsigned, it errors with:

the trait bound `u64: Signed` is not satisfied
the following other types implement trait `Signed`:
  f32
  f64
  i128
  i16
  i32
  i64
  i8
  isize

And the Sized trait cannot be removed since it produce other errors for IntTensorOps.

I also get this error for all usage of execute_with_int_dtype and I don't understand it because the conversion between NdArrayTensor<I> to NdArrayTensorInt is implemented.

error[E0308]: mismatched types
   --> crates\burn-ndarray\src\tensor.rs:312:14
    |
311 |           match ($lhs, $rhs) {
    |                 ------------ this expression has type `(tensor::NdArrayTensor<I>, tensor::NdArrayTensor<I>)`
312 |               ($crate::NdArrayTensorInt::I64(lhs), $crate::NdArrayTensorInt::I64(rhs)) => {
    |                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ expected `NdArrayTensor<I>`, found `NdArrayTensorInt`
    |
   ::: crates\burn-ndarray\src\ops\int_tensor.rs:74:9
    |
74  | /         execute_with_int_dtype!((tensor, source), |tensor, source| {
75  | |             NdArrayMathOps::mask_where(tensor, mask, source)
76  | |         })
    | |__________- in this macro invocation
    |
    = note: expected struct `tensor::NdArrayTensor<I>`
                 found enum `tensor::NdArrayTensorInt`
    = note: this error originates in the macro `execute_with_int_dtype` (in Nightly builds, run with -Z macro-backtrace for more info)

I also noticed some inconsistencies in the file naming conventions between the backends. For example, in the ndarray backend it is named int_tensor.rs, whereas in cubecl it's int_ops.rs.
What about unifying the naming?

Testing

Do I need to run tests for each backend, and where?

[A2va]

Is this right that dtype is not used here ?

[A2va]

A macro might be useful in this case, but I don't know anything about them.

[laggui]

Sorry for the late response! Been a little busy recently 😅 this was on the roadmap so I will absolutely have a look when I get a bit of time

[github-actions]

This PR has been marked as stale because it has not been updated for over a month

[A2va]

@laggui Any update ?