NDArray memory is only released for ScalarNdarray, not any subclass of Ndarray

[BernardoCovas]

Describe the bug

ScalarNdarray implements __del__() to allow python to get rid of the allocated memory, but other subclasses of Ndarray (like VectorNdarray() or MatrixNdarray()) don't implement the same behaviour. Seems to me the __del__() method should be moved to the Ndarray class. I tested this simple change and indeed it fixes the memory leak.

To Reproduce

import taichi as ti
import taichi.types as tt

ti.init(ti.gpu)

while True:
    arr = ti.ndarray(tt.vector(3, ti.f32), (1000, 1000))
    arr = arr.to_numpy()
    # whatch gpu memory skyrocket

If, instead:

import taichi as ti
import taichi.types as tt

ti.init(ti.gpu)

while True:
    arr = ti.ndarray(ti.f32, (1000, 1000, 3))
    arr = arr.to_numpy()
    # whatch gpu memory basically stay the same

Log/Screenshots

Additional comments