Fix for each-blocks preventing outros from completing

[jacwright]

Unkeyed each blocks end up with trailing null values that prevent the whole from being outroed. This fixes it so the null values are removed before outroing the remaining blocks.

Fixes #1617

[Rich-Harris]

Ah, good sleuthing. It feels like we should be resetting the length at the end of the update in the parent to the each-block:

		p: function update(changed, ctx) {
			if (changed.cats) {
				each_value = ctx.cats;

				for (var i = 0; i < each_value.length; i += 1) {
					const child_ctx = get_each_context(ctx, each_value, i);

					if (each_blocks[i]) {
						each_blocks[i].p(changed, child_ctx);
					} else {
						each_blocks[i] = create_each_block(component, child_ctx);
						each_blocks[i].c();
					}
					each_blocks[i].i(each_anchor.parentNode, each_anchor);
				}

				groupOutros();
				for (; i < each_blocks.length; i += 1) outroBlock(i, 1);
+				each_blocks.length - each_value.length;
			}
		},

I'd look into that but I'm about to call it a night — can revisit in the morning if you don't get there first. The advantage of doing it that way is we're not allocating an extra array (and leaving the old one out for the garbage collector).

[jacwright]

I tried that at first, but the aborted-outro-in-each test failed. We can't count the node out until it is finished with its outro.

The most correct way to do it would probably be to splice the node out of the array (instead of setting it to null) when the outro is finished, but I thought the approach I went with would be more performant. I did not do any testing to verify performance behavior between the two.

[Rich-Harris]

Ah, gotcha. Yeah, that's a tough one — splicing versus copying. Probably depends on the specifics and only amounts to a few microseconds. In it goes 😀