[P1] Handle only loop body acceleration in neura

[ShangkunLi]

In the current neura dialect, we assume the kernel is always with loop-related control to ensure a secure start or exit.

However, we may also need to handle the following case:

affine.for (<start, step, bound>) {
  neura.kernel(<ins>) {
    <loop body>
  }
}

where <loop body> is pure computation ops, it is triggered by the outer counter.