Minor FPS gains

[nstbayless]

summary (measured on cv2: belmont's revenge):

• ~5 fps gain from special casing PC < 0x8000 for mch_memory_fetch_PC_2
• ~1 fps gain from using arithmetic rather than double-indirect for accessing rom
• ~1 fps gain from pretending jr $FE costs an extra 20 cycles
• ~ 2 fps gain from using armclang instead of gcc 9.3 (not part of PR)

17 -> 26 fps in opening screen of cloud castle.

Also, I (hopefully) improved the performance of the timer. But cv2 doesn't use the timer, so I haven't checked.

I suggest squash-rebase