Defer clearing tlb flush flag

[ludfjig]

Defer clearing TLB flush flag until after run to prevent losing it if the guest is cancelled before executing the flush. This can happen on windows for example. I noticed this on AMD cpus where guest would sometimes read old/stale memory.

Note: if run returns Ok, the guest hit HLT, meaning it ran through the stub which checks ZF and did the flush. If run returns Err the flush may not have happened but doesn't matter since caller is required to store the sandbox before next dispatch anyway (sandbox is poisoned), which will set the tlb flush flag to true again.

[Copilot]

Pull request overview

Defers clearing the VM’s pending TLB flush flag until after run() completes, aiming to avoid losing the flush request when the guest is cancelled before executing the flush stub.

Changes:

• Stop clearing pending_tlb_flush when setting ZF in RFLAGS.
• Capture run() result and clear pending_tlb_flush after run() returns.