This fixes richgel999#17 but goes further, since it provides higher accuracy for other blocks with few values. Two value blocks are special-cased to use the two endpoints. An early out is taken when the error reaches zero.

As 16-bit we couldn't accumulate the worst-case error without overflowing. Also fixed a bug whereby the values6 were truncated to 8-bit, therefore mostly favouring values8. The return from encode_bc4_hq() is now scaled to the same range from before the changes.