[Feature Request] Fast prefix sum implementation

[RplusW]

Concisely describe the proposed feature
Prefix Sum is used extensively for many particle simulation applications. Currently, the taichi implementation in taichi.algorithm is not extensively optimized compared to the Cupy implementation using CUB backend. I did a side by side comparison of cupy.cumsum with the PrefixSumExecutor and noticed that for size 10 million array, cupy takes 300us while taichi takes 550us. It would be beneficial to add a more optimized PrefixSum algorithm (potentially with allocation of output).

Additional comments
Here are the code I used for benchmarking.

import cupy as cp

test_array = cp.ones((10000000), dtype=cp.int32)

def test_cumsum(test_array):   
    out=cp.cumsum(test_array, axis=None, dtype=cp.int32)

print(benchmark(test_cumsum, (test_array,), n_repeat=100))

import taichi as ti
from taichi.algorithms._algorithms import PrefixSumExecutor

ti.init(arch=ti.gpu, kernel_profiler=True)

array_size = 10000000
array = ti.field(dtype=ti.i32, shape=array_size)
array.fill(1)
PrefixSum = PrefixSumExecutor(array_size)

for i in range(10000):
    PrefixSum.run(array)

ti.profiler.print_kernel_profiler_info()