[aot] A ti::Kernel should support inputs of different data types.

[YuCrazing]

Concisely describe the proposed feature

As mentioned in #6432 , when exporting a kernel with multiple data types, the behaviour of ti::Kernel on the C++ side is unexpected. The reason is that the ti::Kernel does not support inputs of different data types.

Since a kernel on the python side is able to take different data types, the ti::Kernel on the C++ side should have the same behaviour.

Thus, an expected behaviour should be:

python

import taichi as ti

ti.init(arch=ti.vulkan)

@ti.kernel
def k_test(arr: ti.types.ndarray(field_dim=1)):
    arr[2] = arr[0] + arr[1]


a = ti.ndarray(ti.f32, 10)
b = ti.ndarray(ti.i32, 10)

mod = ti.aot.Module(ti.vulkan)

# Serialize a function instance: `k_test_f32`
mod.add_kernel(k_test,
                template_args={
                    "arr": a,
                })

# Serialize a function instance: `k_test_i32`
mod.add_kernel(k_test,
                template_args={
                    "arr": b,
                })

mod.save("aot_files/", "")

c++

// ...

auto a = allocate_ndarray<float>(...);
auto b = allocate_ndarray<int>(...);

ti::Kernel k_test = load_kernel("k_test");

k_test[0] = a;
k_test.launch(); // This should call function `k_test_f32`.

k_test[0] = b;
k_test.launch(); // This should call function `k_test_i32`.

cc @ailzhang