Fix String#scan behavior same as Ruby

[makenowjust]

For example, Ruby's String#scan is:

"hello world".scan(/\w+|(?= )/) # => ["hello", "", "world"]

But Crystal's is:

"hello world".scan(/\w+|(?= )/).map &.[0] # => ["hello", ""]

This pull request fixes it by continuing to scan when match is empty.

[bcardiff]

The change seems good.

I wanted to see if fixing this could cause some inconsistency.
This PR changes thing introduced in 20c8570 . From there I though that checking the behavior of String#split was a good idea. The change from that method has changed in master since 20c857 so the same patch does not longer work.

Currently

In Ruby

"hello world".split(/\w+|(?= )/) # => ["", " "]

In Crystal

"hello world".split(/\w+|(?= )/) # => ["", "", " ", ""]

So, split might be worth of changing as well. Since one is the dual of the other I would vote for changing them at the same time.

[makenowjust]

@bcardiff I fixed String#split behavior, now it works like Ruby and other languages (i.e. JavaScript). But I didn't implement positive and negative limit feature in Ruby because I am always confusing this behavior. (And I specify -1 to limit every time.) Reference

[bcardiff]

The split sample:

assert { "hello world".split(/\w+|(?= )/).should eq(["", " ", ""]) }

differs from Ruby but matches Javascript. I like that result more than the one I extract from Ruby in #3877 (comment).

LGTM