D3D9: Port flush heuristic from D3D11

[doitsujin]

Needs testing.

The flush heuristic itself already works well for D3D11 in practice, with major benefits being that:

• There is no race condition with the pending submission count, unlike in the old code
• We can flush at CS chunk boundaries, even during extremely long command sequences where we otherwise wouldn't
• CPU timing is no longer used as a source of information, which makes this a lot more consistent
• Resource readback latency in general should be much lower.

[misyltoad]

We probably don't want to flush halfway through some commands like StretchRect, ProcessVertices, LockImage and Clear which have a couple of EmitCs'es...

A few of these can probably be changed to be a lot smarter/less shit than the code I wrote like 3 years ago to avoid that too.

[Riesi]

This currently breaks Gothic 3, which wont continue after finishing loading a game. Not sure if Josh's remarks would fix it.

[doitsujin]

Gothic should be fixed.

[Riesi]

yep works now 🍻

[misyltoad]

Looks fine but I still want to merge some of these EmitCses at some point..,