Implemented magnitude

[RianGoossens]

This is AFAIK the first time an implementation turns values of a vectorization factor X to an output of vectorization factor 1. This will be useful for many more functions like .mean(), so I first want to make this pull request before I start making a multitude of new functions.

The way I implemented it is through a new expand function called fixed_output_unary_expand with following signature:

pub fn fixed_output_unary_expand<F>(
    context: &mut CubeContext,
    input: ExpandElement,
    out_item: Item,
    func: F,
) -> ExpandElement

It works identical to unary_expand but it can take in a different output item, could be a different vectorization, could also be a different type, from the input. In principal this function is strictly more general than unary_expand, so perhaps unary_expand should just turn into this function. We can use the same concept to implement new binary functions like dot which also have a different vectorization factor for the output.