[Lang] variable definition, static-typed taichi vs dynamic-typed python

Concisely describe the proposed feature Consider the following codes:

x = ti.var(ti.f32, 4)
x = [0.1, 0.2, 0.3, 0.4] # all below 1.0

@ti.kernel
def hello() -> ti.f32:
  ret = 0  # ret is ti.i32
  for i in x:
    ret += x[i]  # ti.i32 + ti.f32, but still stored in ti.i32!!
  return ret

print(hello())  # user confused: why I got 0.0?

The key reason is that: taichi is statically typed, python is dynamically typed. And the user cost a huge time on debug to realize the problem, then blame us for the inconsistent behavior between python and taichi.

Also consider:

x = ti.var(ti.f64, 4)

@ti.kernel
def hello() -> ti.f64:
  ret = 0.0  # you didn't change default_fp, so it's ti.f32!!
  for i in x:
    ret += x[i]  # ti.f32 + ti.f64, but still stored in ti.f32!!
  return ret # we are losing precision now

Currently we have to depend ti.cast to ensure everything typed correct.

So I may suggest:

1. use x = ti.var(ti.i32) to initialize local variables instead of x = 0.
2. use x = ti.var(ti.f32) to initialize local variables instead of x = 0.0.
3. use x = ti.Vector(2, dt=ti.f32) to initialize local vectors instead of x = ti.Vector([0.0, 0.0]).

Then, we can change the previous code into:

x = ti.var(ti.f32, 4)
x = [0.1, 0.2, 0.3, 0.4] # all below 1.0

@ti.kernel
def hello() -> x.dtype:
  ret = ti.var(x.dtype) # also helps meta programming, right?
  for i in x:
    ret += x[i]
  return ret

print(hello())  # got 1.0, user happy

---

Also consider I want to initialize a vector y, whose dimension and type are same with x (meta programming). Currently, we can only:

y = ti.Vector([(0.0 if ti.core.is_real(x.dtype) else 0) for i in range(x.shape[0])])

Or use a dirty hack:

y = x
y *= 0

But after this issue, we can:

y = ti.Vector(x.shape[0], dt=x.dtype)`

more clear, more easy, right?

Describe the solution you'd like (if any) Not yet have a sol.

Additional comments What will ti.Vector([0.0, 0]) do?