Fix race on reading intrabc blocks in above wavefront rows

[mrt-amd]

Adds extra synchronisation delay to ensure that all above blocks which are valid targets for global intrabc are already available when the current block encode runs.

Fixes #5571.

This can decrease speed noticeably on row-mt intra-only encodes with global intrabc enabled as row threads don't parallelise as well (modes not using it, including inter frames, are not affected).

I think the current proposal is the best solution, though there are alternatives with different tradeoffs:

• Disallow global intrabc vectors from pointing into the unavailable region. This is directly a small quality loss.
• Synchronise only if required. Wait at the point where the encoder wants to look at a vector into a not-yet-coded block, which would be either when the hash search finds a match there or when the regular search enter that region. Probably this happens less and the impact might therefore be lower, but implementation looks a lot more complex.