Certain pointer-based SNode tree behaves incorrect in packed mode

[g1n0st]

Describe the bug

When packed=True in global config and the SNode Tree is pointer(3, 3)-dense(8, 8), struct-for loop will omit certain index even it has been activated:

To Reproduce

import taichi as ti
ti.init(arch=ti.cpu, packed=True)
​
block_size, leaf_size = 3, 8
​
x = ti.field(ti.i32)
block = ti.root.pointer(ti.ij, (block_size, block_size))
block.dense(ti.ij, (leaf_size, leaf_size)).place(x)
​
print(x.to_numpy().shape)
​
@ti.kernel
def activate():
  x[7, 7] = 1
​
activate()
​
@ti.kernel
def test():
  for I in ti.grouped(x):
    if I[0] == 7 and I[1] == 7: print(I)
​
test()
​
# Expected Output:
# (24, 24)
# [7, 7]
​
# Actual Output (iff block_size == 3 and x[7, 7] = 1):
# (24, 24)