Fix float range behavior

[simonster]

Previously:

1.0:1.5      # == 1.0:2.0
1.0:1.0:1.5  # == 1.0:1.0

With this patch:

1.0:1.5      # == 1.0:1.0
1.0:1.0:1.5  # == 1.0:1.0

I don't think the previous behavior was intentional, since it seems to violate the principle of least surprise, but I could be wrong.

[JeffBezanson]

If we want x:y and x:1.0:y to always be equivalent, I believe the only option is to define

colon{T<:FloatingPoint}(start::T, stop::T) = colon(start, 1.0, stop)

But it also makes sense to insist on a Range1 from 2-argument colon, in which case your change is fine.

[simonster]

I'd be fine with returning a Range, but if we want to return a Range1, maybe we could do something like:

function colon{T<:Real}(start::T, stop::T)
    if stop < start
        len = 0
    else
        nf = stop - start + 1
        if T <: FloatingPoint
            n = round(nf)
            len = abs(n-nf) < eps(n)*3 ? itrunc(n) : itrunc(nf)
        else
            n = nf
            len = iround(n)
        end
        if n >= typemax(Int)
            error("Range: length ",n," is too large")
        end
    end
    Range1(start, len)
end

Not exactly the same as x:1.0:y, but more intuitive than what's currently in this PR for cases like 1.0:(.3-.1)/.1

[JeffBezanson]

Let's go with that, except the last len = iround(n) should be len = itrunc(n). That change is also needed in the existing 3-argument colon.

[simonster]

Done.

[JeffBezanson]

Great, thanks!