Support dataflow primitive

[hecmay]

Referred to Hongzheng's PR. We need to handle two cases.

1. dataflow optimization inside function body

# Apply .to to place data
s.to(kernel.conv, kernel.out, depth=1)
s.to(image, kernel.out, depth=20)

# Mark function body as dataflow region
top = s.subgraph()[0]
s[top].dataflow()

2. dataflow optimization inside a loop

loop = hcl.compute((10,32), lambda *args: pe(args), "loop")

# Mark 1st loop's body as dataflow region
s[loop].dataflow(axis=1)

[zhangzhiru]

Which one is not supported?

[hecmay]

@zhangzhiru. 1st one.