freedreno: GMEM layout improvements
I noticed that compared to blob on manhattan, we were loosing ground on the 4x MRT pass, which I traced down to a suboptimal GMEM layout. This MR re-works the GMEM layout algo, which picks up a bit more than 3-4% on mh.. and 4-5% on trex.
(also an unrelated patch to make it easier to enable fdlog on android.. since it is silly to keep carrying this around privately but no point in giving it it's own CI runs)