Don't compute block_def_map when there are only macros in expression position

[jonas-schievink]

Currently we cannot distinguish between macro calls in item/statement vs. expression position, leading to unnecessary block_def_map computations. For example, in this code there's no need to compute block_def_map for the function block:

fn f() {
    let x: mac![] = e!();
}

This is a very minor issue and is unlikely to impact performance or memory usage much.

[lnicola]

Do we still want to do this? Not that I think either way, but we're lagging on ungrammar.

[jonas-schievink]

Crap, I shouldn't have published that so eagerly, I ran into some issues while updating r-a

[jonas-schievink]

I reverted the ungrammar changes since I came up with a simpler way to achieve this: We "just" compute per-block ItemTrees and skip block_def_map computation when that ItemTree has no top-level items.