Smarter @eachrow

[pdeffebach]

As noted in #248, we currently always make a full copy of df in every @eachrow. Users perform the code

s = let x = 0
    @eachrow df begin 
        x += 1
    end
    x
end

and expect fast code, when in actuality we are creating a whole new data frame and there will be tons of allocations.

I'm not sure there's a lot we can do. It makes me wonder if we should make the ! as the default for everything so users always get the fastest code.

[bkamins]

I assume @eachrow can mutate the copy of df. Right?

In this case I think it is OK that it makes a copy. However, I think it would be good to clearly highlight in the documentation in what cases what solution should be used.

[pdeffebach]

Yes. I will emphasize the copying aspect and discuss performance in docstrings in #248

[nalimilan]

Yeah the behavior is potentially confusing but it makes sense in general. The only case when it's weird is when you don't mutate columns but only accumulate into a global variable.