Fix HXCPP_GC_BIG_BLOCKS row count regression

[dimensionscape]

#1402 changed mUsedRows and mFraggedRows in BlockDataInfo from int to uint8_t. That's enough for the default 32k blocks, which have 254 usable lines, but HXCPP_GC_BIG_BLOCKS uses 64k blocks with 508 usable lines, so the counts wrap once a block has more than 255 lines in use.

The worst effect is that the GC badly underestimates how much memory is in use after a collection. The working-memory target it sets stays far too small, so it ends up collecting almost every time it needs a new block, and a big-blocks build with a reasonable amount of live data spends most of its time in the GC. The wrap also means countRows() never sees a full block as full, and a block with exactly 256 used lines looks empty.

This changes both fields back to uint16_t, along with the allocator's mFraggedRows pointer and the casts into it. It also adds static_asserts so the counts can't end up smaller than IMMIX_USEFUL_LINES again. The byte counts are fine as uint16_t (508 x 128 = 65024).

See Repro:

Build with -D HXCPP_GC_BIG_BLOCKS and compare against a build without it. With the define, the churn loop is dramatically slower.

class Main {
	static function main() {
		var t0 = haxe.Timer.stamp();
		var live = new Array<Node>();
		for (i in 0...6000000)
			live.push(new Node(i));
		var t1 = haxe.Timer.stamp();
		var sum = 0;
		for (i in 0...30000000)
			sum += new Node(i).v & 1;
		var t2 = haxe.Timer.stamp();
		Sys.println('live set ${Std.int((t1 - t0) * 1000)}ms, churn ${Std.int((t2 - t1) * 1000)}ms ($sum, ${live.length})');
	}
}

class Node {
	public var v:Int;
	public var next:Node;

	public function new(v:Int)
		this.v = v;
}

Note: on current master you may also hit the marking crash fixed by #1407 while testing this.

[tobil4sk]

This comment doesn't seem very useful to keep, since it refers to the uint8_t that is removed by this PR. I think the static assert is more helpful at documenting the requirement

[dimensionscape]

Yeah, that's fair. It was originally a warning against regression, but the static asserts keep us safe. I'll remove it.