Autotune: support int/bool ops

[louisfd]

In autotune, we create random tensors as inputs. Since random only works with floats it makes it impossible to use autotune with int and bool.

Lazy solution: revert to creating ones tensor instead of random
Hacky solution: Run the random kernel on f32 instead of E then cast to int or bool using some
Long solution: Add support at large for bool random (bernoulli) and int random (uniform)

Once this is done we can add a call to autotune for sum_dim and mean_dim (int versions)

[agelas]

@antimora I think I'll have some more free time coming up, I can tentatively take this.