[WIP] float_dtypes(device) -> Vec<FloatDType> sketch.

[crutcher]

This is a first-pass sketch at threading introspection operations for building a B::Device => capability mapping.

This is mostly trivial, but I don't know what to do about the Router case.

[nathanielsimard]

I think it would be better to have a B::is_dtype_supported(device, dtype) -> bool. It avoids an allocation and it assigns the dtype to the device. A backend can have multiple devices with different dtype support.

[laggui]

[cross-posting my earlier discord comment here]

Hmmm I don't think this should be defined for the FloatTensorOps trait. Also, in line with the original tracking issue, I think we should have something like:

let is_supported = B::supports_dtype(&device, dtype);

And if we really want a matrix of supported dtypes, would need something like

pub struct BackendDTypes {
    pub float: Vec<DType>,
    pub int: Vec<DType>,
    pub bool: Vec<DType>,
}

fn supported_dtypes(&device) -> BackendDTypes;

but not sure if that is desirable.