Support non-consuming operations

[kyleheadley]

I'm calculating both mean and stddev on some data, but I find that I have to clone my data vec to do it. The functions as they are consume an iterator, whose item type must implement ToPrimitive. But ToPrimitive is not defined over &i32 and the like, so I can't use, for example, mean(vec![1,2,3].iter()). I had to use something like mean(vec![1,2,3].clone().into_iter(). I had small examples, but I think you want this library to scale, and cloning data to calculate stats won't scale.

[BurntSushi]

This should work:

let xs = vec![1, 2, 3];
let avg = mean(xs.iter().cloned());
println!("xs is still mine: {:?}", xs);

The trick here is that xs.iter() provides an Iterator<Item=&i32> and cloned turns the &i32 into an i32, which should be optimized out.

[kyleheadley]

Ok, that optimization should work in the common cases. However, it's still cloning before converting to an f64, rather than converting from a reference to the data. You'll have to suggest to people to impl ToPrimitive for &'a Bigdata if for some reason they have a Bigdata type that converts to a primitive and want stats.

[BurntSushi]

Yes, if your type doesn't impl Copy then your suggestion sounds good. I think all of the types that impl ToPrimitive in the num_traits crate impl Copy except for BigInt and BigUint.

More broadly: I don't really have a good vision for this crate. My hope is that something more cohesive takes its place, but I'm not really the right person to lead that charge. I should probably put that in the README.