sh4: end the current timeslice on Stop()

[Immersion95]

This PR supersedes #2309 with a cheaper implementation of the same fix.

The previous PR fixed the issue by making rec-x64 observe CpuRunning between translated blocks, but that added work in the hottest part of the x64 dynarec.

This version takes a lower-cost approach instead:

• Stop() now also ends the current SH4 timeslice immediately
• rec-x64 exits before UpdateSystem_INTC() if CpuRunning is already false

In practice, this keeps the same useful behavior in the affected cases: execution stops before the following render starts, which restores the expected delayed frame swap behavior in single-threaded Win64 GL/DX11 and also fixes the PenPen loading screen cut.

So compared to #2309, the main benefit is that this avoids a per-block CpuRunning check in the hottest rec-x64 path while preserving the fix.

Fixes #1615
Fixes #2308

[flyinghead]

This is much better. A few comments:

• Please target the dev branch instead of master (this will also fix the failing CI jobs)
• You can put the CpuRunning test after all cycle_counter processing has been done, so you won't need to reload rax:

  		test(ecx, ecx);
		jg(slice_loop);

		add(ecx, SH4_TIMESLICE);
		mov(dword[rax], ecx);

+		mov(rax, (size_t)&sh4ctx.CpuRunning);
+		mov(edx, dword[rax]);
+		test(edx, edx);
+		je(end_run_loop);

		call(UpdateSystem_INTC);
		jmp(run_loop);

• Note that after UpdateSystem_INTC returns, it jumps to run_loop, and starts by testing CpuRunning. So I wonder if the newly added CpuRunning test is really needed. I don't think calling UpdateSystem_INTC before stopping would be a problem but I haven't tested it.

[Immersion95]

You are totally right, that's an even simpler fix :).

I tested that simpler fix based on your comment, and the extra CpuRunning check in rec-x64 is not needed. Just clearing cycle_counter in Stop() is enough to fix the affected cases on my side.

I've simplified the PR to only change Stop(), and I’ve retargeted it to dev :)

[flyinghead]

Nice