freedreno: Stop tracking reloc flags to reduce draw overhead

append_bo() was still a terrifying amount of our draw overhead, so let's try to reduce that by not passing reloc flags through the stack. This matches how Intel and I believe AMD and other drivers behave with dma-bufs, and deletes a bunch of code and an easy way to make hard-to-find mistakes.

Once this is in place, I'm hoping to use a different datastructure (like ANV does) to track BOs being in the relocs list already.

Depends on !4957 (merged)