Improved template loop generation

[vallentin]

Fixes #107
Fixes #333
Closes #221

No test cases were modified, and this is a non-breaking change.

This PR improves the generation of TemplateLoop, such that it has a higher chance to produce Rust code that will actually compile, based on assumptions that can be drawn from the context of the expression used. See the comments in the code, for what those assumptions are.

---

Old Explanation (from #396 related to this change):

Based on the context of where it is used, what can it be assumed to be?

• With loops the assumption can be made that:
  • *Call(..) and Index(..) likely return something that implements IntoIterator so don't borrow
  • If not a *Call(..) but self. is used, then assume it needs to be borrowed
  • Otherwise, assume that IntoIterator is implemented and the value can be moved

[vallentin]

Forgot to include "(fixes #107, #333)" in the commit message, guess you can add it while merging :)

[djc]

For this change, it would actually be good just to add some simple test cases that exercise the new heuristics, for example, one test case per match arm you've added. Sound good?

I also prefer to rebase PRs, in which case I don't get to rewrite the commit message, so please just add the issue numbers. Should be easy with git commit --amend && git push -f.

[vallentin]

I've added a few tests.

---

I wanted to add a test using repeat().take() and/or zip(), but that won't work until I've made a PR (again) for is_copyable and not-borrowing literals. However, before I do that, I'll look into it a bit more, as I think there's a few cases that would be non-breaking changes.

#[derive(Template)]
#[template(
    source = "{% for i in ::std::iter::repeat(\"a\").take(5) %}{{ i }}{% endfor %}",
    ext = "txt"
)]
struct ForPathCallTemplate {}

#[test]
fn test_for_path_call() {
    let t = ForPathCallTemplate {};
    assert_eq!(t.render().unwrap(), "aaaaa");
}

[djc]

This is awesome, thanks!