[opt] Add more accurate alias analysis for ExternalPtrStmt

[strongoier]

This PR is a follow-up of #2952. Making alias analysis more accurate can unblock future optimizations on ExternalPtrStmt (many of current optimization passes only focus on GlobalPtrStmt).

That said, we can already observe some benefits of this PR. For example, regarding the following code snippet,

@ti.kernel
def test(a: ti.any_arr(element_dim=1)):
    for i in range(10):
        a[i][0] = i
        a[i][1] = i * 2
        a[i][2] = a[i][0]

a = np.zeros((10, 3))
test(a)

Before this PR, the final CHI IR is,

kernel {
  $0 = offloaded range_for(0, 10) grid_dim=0 block_dim=32 
  body {
    <i32> $1 = loop $0 index 0
    <*f64> $2 = arg[0]
    <i32> $3 = const [0]
    <*f64> $4 = external_ptr <$2>, [$1, $3]
    <f64> $5 = cast_value<f64> $1
    $6 : global store [$4 <- $5]
    <i32> $7 = const [1]
    <i32> $8 = bit_shl $1 $7
    <*f64> $9 = external_ptr <$2>, [$1, $7]
    <f64> $10 = cast_value<f64> $8
    $11 : global store [$9 <- $10]
    <f64> $12 = global load $4
    <i32> $13 = const [2]
    <*f64> $14 = external_ptr <$2>, [$1, $13]
    $15 : global store [$14 <- $12]
  }
}

After this PR, the final CHI IR is,

kernel {
  $0 = offloaded range_for(0, 10) grid_dim=0 block_dim=32 
  body {
    <i32> $1 = loop $0 index 0
    <*f64> $2 = arg[0]
    <i32> $3 = const [0]
    <*f64> $4 = external_ptr <$2>, [$1, $3]
    <f64> $5 = cast_value<f64> $1
    $6 : global store [$4 <- $5]
    <i32> $7 = const [1]
    <i32> $8 = bit_shl $1 $7
    <*f64> $9 = external_ptr <$2>, [$1, $7]
    <f64> $10 = cast_value<f64> $8
    $11 : global store [$9 <- $10]
    <i32> $12 = const [2]
    <*f64> $13 = external_ptr <$2>, [$1, $12]
    $14 : global store [$13 <- $5]
  }
}

We can see that a redundant global load can now be eliminated.

[netlify]

✔️ Deploy Preview for jovial-fermat-aa59dc canceled.

🔨 Explore the source changes: b69a4c9

🔍 Inspect the deploy log: https://app.netlify.com/sites/jovial-fermat-aa59dc/deploys/61c4043a69cac20007f33132

[ailzhang]

Awesome!! This might help ndarray perf as well!