[P1] grant_always might be needed

[tancheng]

We might also need a neura.grant_always to distinguish from neura.grant_once for the for loop when there is a live-in from the entry block https://github.com/coredac/dataflow/blob/main/test/neura/ctrl/branch_for.mlir.

[tancheng]

Initial thought is

• Values (existing in entry block) passed in other blocks as directly live-in (not via block argument) should be grant_always, as they should be forever/always valid
• On the other hand, values passed in as block argument (via a br/jump or cond_br) should be grant_once

Sounds like a simple fix, not sure whether I missed anything. CCing @Yfeng-44 and @ShangkunLi for visibility.