Add macro to automatically generate common cases for autotune

[wingertge]

Adds a small proc macro to generate the autotune boilerplate in most common cases.
See examples in doc comments for how they would be used.

[nathanielsimard]

Just to make sure, this creates the tune operation as well as the autotune set from an autotune key, but it doesn't create the high level function actually creating a static tuner right?

[wingertge]

Just to make sure, this creates the tune operation as well as the autotune set from an autotune key, but it doesn't create the high level function actually creating a static tuner right?

Exactly. That boilerplate is extremely minor so I don't think it's worth making a macro for.
Here's an example of the usage in the new conv2d tuner

[nathanielsimard]

Exactly. That boilerplate is extremely minor so I don't think it's worth making a macro for. Here's an example of the usage in the new conv2d tuner

We could actually generate it as well, since it's the same signature as the tune_set function.

We could have another macro tune that does most of the thing:

#[tune(
    operations(Conv2dDirect, Conv2dIm2col),
    key = create_key,
    inputs = bench_inputs
)]
pub fn conv2d<R: JitRuntime, E: FloatElement, I: IntElement>(
    key: JitAutotuneKey,
    input: JitTensor<R, E, 4>,
    weights: JitTensor<R, E, 4>,
    bias: Option<JitTensor<R, E, 1>>,
    options: ConvOptions<2>,
) -> JitTensor<R, E, 4>;

What do you think?