Problem compiling CUDA code with new const

[torsteingrindvik]

In f81a2fa we got const T * __restrict__ for readonly inputs which is great 🎉

However it seems some patterns won't compile with this:

thread 'main' panicked at cubecl/crates/cubecl-cuda/src/compute/server.rs:380:17:
[Compilation Error]
    default_program(77): error: a value of type "const float *" cannot be used to initialize an entity of type "float *"
      float *slice_2_0 = input_2 + l_0_17;

So input_2 is const float* __restrict__ input_2,

and it's being used like this: float *slice_2_0 = input_2 + l_0_17;

I think the relevant kernel code is something like:

let slice = self.data.slice(start, start + foo);

So I think perhaps there is a missing const in the cpp source when trying to slice const inputs.

[nathanielsimard]

yeah we removed it because of that, I added a TODO in the code to add it back when we fix the slice notation to consider const __restrict__.