Split TryConcatAlong into different traits

[swfsql]

Closes #891.

• Added a failing test.
  • Unsure where to add the test. It's about a Module, so it should not be in dfdx_core. But there's no Module in dfdx::nn representing tensor concatenation, so I've added as an integration test.
  • The test was later adjusted to the new trait, being able to run successfully.
• Deprecated TryConcatAlong in favor of TryConcatTensorAlong or TryConcatShapeAlong.
• Created concat_tensor_along/ and concat_shape_along/.
  • Copied relevant sections and files from concat_along, adjusting where necessary.
  • Moved concat_along/ kernels to concat_tensor_along/.

The added test, that initially failed to compile:

dfdx/dfdx/tests/issue_tests.rs

Lines 4 to 44 in b292254

	#[test]
	fn test_issue_891() {
	#[derive(Default, Debug, Clone, Copy, CustomModule)]
	pub struct Id;
	impl<Input> Module<Input> for Id {
	type Output = Input;
	fn try_forward(&self, x: Input) -> Result<Self::Output, Error> {
	Ok(x)
	}
	}
	#[derive(Default, Debug, Clone, Copy, dfdx_derives::CustomModule)]
	struct ConcatTensorAlong<Ax: Axes<Array = [isize; 1]> + Debug>(pub Ax);
	impl<Input, const AXIS: isize> Module<Input> for ConcatTensorAlong<Axis<AXIS>>
	where
	Input: TryConcatAlong<Axis<AXIS>>,
	{
	type Output = <Input as TryConcatAlong<Axis<AXIS>>>::Output;
	fn try_forward(&self, x: Input) -> Result<Self::Output, Error> {
	x.try_concat_along(Axis)
	}
	}
	type Arch = (SplitInto<(Id, Id)>, ConcatTensorAlong<Axis<0>>);
	let dev = Cpu::default();
	let x = dev.tensor([1.]);
	let m = dev.build_module::<f32>(Arch::default());
	let y = m.forward(x);
	/*
	error[E0275]: overflow evaluating the requirement `((_, _, _, _), (..., ..., ..., ...)): dfdx::prelude::TryConcatAlong<...>`
	--> dfdx/tests/issue_tests.rs:36:15
	|
	36 | let y = m.forward(x);
	| ^^^^^^^
	|
	*/
	}

[coreylowman]

Some tiny changes, but otherwise looks good and I was able to verify that the current main doesn't compile the test and the fix does 👍

[coreylowman]

@swfsql looks good - do you mind merging with the webgpu changes I just merged? Should be pretty straightforward 🤞 No need to test with webgpu features yet

[swfsql]

@coreylowman sure np, I've rebased and basically just moved the webgpu kernel to where the others are.

I've made one change, remaking this item as public so it's the same behavior as from before.

dfdx/dfdx-core/src/tensor_ops/concat_along/mod.rs

Line 1 in 2a31bdf

pub use super::concat_tensor_along::ConcatAlongKernel;

[coreylowman]

Woohoo, thanks for this change! 🎉