Investigate correctness of unified memory fix for multi-GPU system

[MaxKev1n]

I attempted to fix the unified memory issue in the latest version of MGPUSim for multi-GPU systems.

In my current implementation, I discard all in-flight translation and memory requests during page migration, and let compute units reissue the memory requests after migration completes.  I modified both Akita and MGPUSim to support this behavior. However, I am not sure whether this approach is functionally correct.

https://github.com/MaxKev1n/akita_public/commit/719422d3016a45ab985a1c235ad795633ef2b417 https://github.com/MaxKev1n/mgpusim_public/commit/37c5a3349ee4f72f91f0478dc54d2e64465e67a0