Enable the safe return in neura dataflow ir

[ShangkunLi]

In this pr:

1. Introduce two kinds of return ops in neura dataflow ir:

   • neura.return_void: its operand means the trigger values
   • neura.return_value: its operand means the returned values
     These two kinds of return ops are only enabled in dataflow ir. In conventional neural CDFG ir, we still use the neura.return.

2. We create a canonicalize-return pass to mark the type (return_type = void/value) of neura.return. And

3. We introduce some logic in the transform-ctrl-to-data-flow pass to transform the neura.return to neura.return_value/return_void

4. A neura.yield is added at each flattened block's end to act like a terminator for the block. This is just to obey the rules in MLIR block https://mlir.llvm.org/docs/LangRef/#blocks.

TODO: Enable the interpreter for these two introduced ops. Since this pr is too large, I will make a separate pr for that.