[ndarray][doc] Improve behaviors of 0darray + more docstring

[forfudan]

This PR aims to improve the behaviors of 0-dimensional array (numojo scalar). Note that a.item(0) or a[Item(0)] is always preferred because the behavior is more determined, but we also allow some basic operations on 0darray to make users' life easier.

0-dimensional array cannot be constructed by users but can be obtained by array indexing and slicing. Printing this variable gives the scalar and a note that it is an 0darray instead of a mojo scalar. It is similar to numpy in that a[0] returns a numpy scalar and a.item(0) returns a Python scalar. For example,

>>> var a = nm.random.arange[f16](0, 3, 0.12)
>>> print(a[1])
0.11999512  (0darray[f16])
>>> print(a.item(1))
0.11999512
>>> var c = nm.array[f16]("[[1,2], [3,4]]")
>>> print(c[1, 1])
4.0  (0darray[f16])

0-dimensional array can be unpacked to get the corresponding mojo scalar either by [] or by item(). For example,

>>> var a = nm.random.arange[f16](0, 3, 0.12)
>>> var b = a[1]
>>> print(b)
0.11999512  (0darray[f16])
>>> print(b[])  # Unpack using []
0.11999512
>>> print(b.item())  # Unpack using item()
0.11999512

0-dimensional array can be used in arithmetic operations, just like a scalar.

>>> var a = nm.random.arange[f16](0, 3, 0.12)
>>> var b = a[1]
>>> var c = nm.array[f16]("[[1,2], [3,4]]")
>>> var d = c[1, 1]
>>> print(b - d)  # Arithmetic operations between two 0darrays
-3.8808594  (0darray[f16])
>>> print(b[] - d[])  # Arithmetic operations after unpacking
-3.8808594
>>> print(b < d)  # Comparison between 0darray and 0darray
True  (0darray[boolean])
>>>print(b == d[])  # Comparison between 0darray and unpacked 0darray
False  (0darray[boolean])