[bug] TensorType arithmetic gives wrong results

[turbo0628]

As a follow-up of #7061, I found some vector arithmetics gave me wrong numerical results:

import taichi as ti

ti.init(arch=ti.cuda, print_kernel_llvm_ir_optimized=True, print_ir=True, real_matrix_scalarize=False)

f = ti.ndarray(ti.math.vec4, 4)

@ti.kernel
def fill(arr: ti.types.ndarray()):
    v0 = ti.math.vec4([0.0, 1.0, 2.0, 3.0])
    v1 = ti.math.vec4([1.0, 2.0, 3.0, 4.0])
    v2 = ti.math.vec4([2.0, 3.0, 4.0, 5.0])
    v3 = ti.math.vec4([4.0, 5.0, 6.0, 7.0])
    arr[0] = v0
    arr[1] = v1
    arr[2] = v2
    arr[3] = v3

@ti.kernel
def vec_test(arr : ti.types.ndarray()):
    v0 = arr[0]
    v1 = arr[1]
    v2 = arr[2]
    v3 = arr[3]

    arr[0] = v0 * v1 + v2
    arr[1] = v1 * v2 + v3
    arr[2] = v0 * v2 + v3

fill(f)
vec_test(f)
print(f.to_numpy())

Screen output was

[[ 2.  6.  4.  9.]
 [16. 31.  7.  0.]
 [ 0.  0.  0.  0.]
 [ 0.  0.  0.  0.]]

Which is apparently unexpected.