"-z" (minddata) and "-M on" (gpu) by default in build.sh

fixed: PrimitiveToInferImplMap map is global, and key of the map PrimitivePtr also a global variable. If key is initialized later than the map initialized during compilation, will cause the primitive map initialize failed. Variable initialization order is not guaranteed during compilation.

num threads cannot be more than num shards

minor fix

clang style fix

address review comments

This commit is to separate the computation cost and memory cost in auto_parallel. Some related memory correction is removed.