serialize upon syscall entry

[wangbj]

This is a follow up of #34. Roughly after commit 93333a6, we make our scheduler schedule one-thread-at-a-time. However, there're several downside (beside PRO being more deterministic):

• performance degration
• live locks, such as Haskell parallel GC use spin locks for synchronization (https://github.com/ghc/ghc/blob/bf73419518ca550e85188616f860961c7e2a336b/rts/sm/GCUtils.c#L37)

It should be a safe assumption a task is deterministic unless it enters syscalls. hence we can allow parallel execution, upon entry of syscalls, which is serialized.