Support lowering linalg.generic to affine.for

[aqjune]

1. when the linalg.generic loop is full parallel & the affine.for contains memref ops
2. when the linalg.generic loop does reduction..? the memref store must put e.g. summation