GPU memory is not released. How to release memory for specific field variable?

[qmdxcube]

Describe the bug
GPU memory is not released.

To Reproduce
Please post a minimal sample code to reproduce the bug.

# taichi-1.5.0
import taichi as ti
import numpy as np

ti.init(arch=ti.cuda,kernel_profiler=True,device_memory_GB=5.0)

nelx, nely, nelz = 256, 256, 192
elements_box = (nelx,nely,nelz)
nodes_box = (nelx+1,nely+1,nelz+1)

# design variable vector and sensitivity
x = ti.field(dtype=ti.f64, shape=elements_box)
xTilde = ti.field(dtype=ti.f64, shape=elements_box)
xPhys = ti.field(dtype=ti.f64, shape=elements_box)
xold = ti.field(dtype=ti.f64, shape=elements_box)
dc = ti.field(dtype=ti.f64, shape=elements_box)
dv = ti.field(dtype=ti.f64, shape=elements_box)
dcPhys = ti.field(dtype=ti.f64, shape=elements_box)
dx = ti.field(dtype=ti.f64, shape=elements_box)

F = ti.Vector.field(3, dtype=ti.f64, shape=nodes_box)
supports = ti.field(dtype=ti.i32, shape=nodes_box)

u = ti.Vector.field(3, dtype=ti.f64, shape=nodes_box)
r = ti.Vector.field(3, dtype=ti.f64, shape=nodes_box)
p = ti.Vector.field(3, dtype=ti.f64, shape=nodes_box)
q = ti.Vector.field(3, dtype=ti.f64, shape=nodes_box)

use_multigrid = True
n_mg_levels = 7

def node_shape(l):
	return (nelx//2**l+1, nely//2**l+1, nelz//2**l+1)

def node_number(l):
	return (nelx//2**l+1)*(nely//2**l+1)*(nelz//2**l+1)

A = [ti.Matrix.field(m=3,n=3,dtype=ti.f64,shape=(node_number(l),27)) for l in range(2,n_mg_levels)] # stencils for coarser grids
M = [ti.Vector.field(3, dtype=ti.f64, shape=node_shape(l)) for l in range(n_mg_levels-1)]  # multigird inverse of diagonal of A
Az = [ti.Vector.field(3,dtype=ti.f64,shape=node_shape(l)) for l in range(n_mg_levels)]
rz = [ti.Vector.field(3, dtype=ti.f64, shape=node_shape(l)) for l in range(n_mg_levels)]  # multigird residual for Az=r
rAz = [ti.Vector.field(3, dtype=ti.f64, shape=node_shape(l)) for l in range(n_mg_levels)]  # multigird residual for Az=r
v = [ti.Vector.field(3, dtype=ti.f64, shape=node_shape(l)) for l in range(n_mg_levels)]  # multigird residual for Az=r
z = [ti.Vector.field(3, dtype=ti.f64, shape=node_shape(l)) for l in range(n_mg_levels)]  # multigird solution for Az=r

if __name__ == '__main__':
	change = 1.0
	while change > 0.01:
		change = 0.005

	del x,xTilde,xold,dc,dv,dcPhys,dx
	del F,supports
	del r,p,q
	del A,M,Az,rz,rAz,v,z

	xPhys_to_show = xPhys.to_numpy()
	U_to_show = u.to_numpy()

Log/Screenshots
My GPU has 6GB memory and 5GB was used for this code. Error appears when the last line of the code is running. Strangely, when the last two lines are exchanged, the code can be run successfully. Some fields are only useful for computation, so they can be deleted after computation and the del function was employed in this code. Other fields will be transformed into numpy arrays for further purpose. What is the reliable way to release memory for specific field variable? Thanks.

![traceback](https://github.com/taichi-dev/taichi/assets/67142522/b2763083-1eec-4cdb-ba67-ab706e0bdfc0)

Additional comments
If possible, please also consider attaching the output of command ti diagnose. This produces the detailed environment information and hopefully helps us diagnose faster.

If you have local commits (e.g. compile fixes before you reproduce the bug), please make sure you first make a PR to fix the build errors and then report the bug.

[lin-hitonami]

You can use ti.FieldsBuider to manually manage the allocation and destruction of fields. See here for more information.

[lin-hitonami]

Closing this issue. If you have further questions, feel free to reopen it.