Replace LinalgScalar by num_traits::Num

[termoshtt]

ndarray::linalg_traits::LinalgScalar and num_traits::Num are basically same trait. num-traits 0.2 has more specific traits i.e. Num, NumOps, NumRef, NumAssign, NumAssignOps and NumAssignRef. LinalgScalar is merely used in this crate, and it exists for user convenience. Should we replace LinalgScalar by Num?

[jturner314]

ndarray::LinalgScalar requires the following: 'static + Copy + Zero + One + Add + Sub + Mul + Div. I think we could probably eliminate the Copy bound, but we can't really get rid of the 'static bound right now. (See rust-ndarray/ndarray-stats#3 for the reasoning for keeping 'static.)

num_traits::Num requires the following: PartialEq + Zero + One + Add + Sub + Mul + Div + Rem.

So, we could replace LinalgScalar bounds with 'static + Copy + Num, or possibly 'static + Num. To be honest, though, I don't really see any reason to do so. LinalgScalar is automatically implemented for T where T meets the required bounds, so it's not an issue of convenience. If we switch to 'static + Copy + Num, we're adding unnecessary Num + Rem bounds (and note that Num has to be implemented for the specific type, unlike LinalgScalar which is implemented for T).

Is there any reason in particular to make this change?

[termoshtt]

Your comment for 'static is quite reasonable.

Is there any reason in particular to make this change?

The generalization of number types is responsible for the num-traits crate. IMO, we should not introduce a trait for fine-tuned minimal requirement (LinalgScalar) rather than using commonly used trait ('static + Num).

If we switch to 'static + Copy + Num, we're adding unnecessary Num + Rem bounds

This is actually problematic...
num-complex::Complex had been lacking these traits long time until rust-num/num#327. Num is not a large problem, but Rem is sometimes hard to implement.

[jturner314]

We could use 'static + Copy + NumOps to avoid the Num requirement, but NumOps still requires Rem. It looks like all of the relevant num-traits traits require Rem.

[termoshtt]

I make sense the difference of LinalgScalar and Num. Thanks.