Fix busy waiting on finish_flag

[Sacusa]

If the program that invokes the accelerator loops on the finish_flag instead of waiting for an interrupt, it will likely panic and die. This is caused by the accelerator writing to an uncacheable memory region and the CPU MSHR being unaware of it. Because the finish_flag is uncacheable, the accelerator's write request sets the needsWriteable flag of the memory packet but not the isInvalidate flag. The CPU MSHR, which is not aware of the presence of uncacheable memory, assumes both must be set or neither, triggering the panic.

[xyzsam]

Can you share an example where this panic occurs? We have not observed it in any of our current workloads (which can get very large). Also, I don't think the accelerator should be writing to an uncacheable region, nor should it need to. The only requirement is that the CPU refer to it as a volatile int so that it's always reloaded from memory.