Convenience of mixed tensor scalar operations

[jackerschott]

Suppose I would want to write something like

fn mul_by_two<S : Shape, E : Dtype, D: Device<E>, T : Tape<E, D>>(x : Tensor<S, E, D, T>)
    -> Tensor<S, E, D, T>
{
    return x * E::from_f32(2.0).unwrap();
}

If I see this correctly I always have to use from_f32 and unwrap for converting floats to construct a specific generic scalar, which makes larger mathematical functions like this hard to read. It might be an idea to implement a from method (with associated try_from) that can be called as just E::from(2.0)? This would break with the rust conventions though, since from usually cannot panic, if I see this correctly.

[coreylowman]

Oooh we could change all the scalar impls to accept Rhs: Into<f64> instead of E. That should address this? Should clean up the f16 impls as well