candle-nonzero

cpu/cuda nonzero op of candle framework

add nonzero op for candle framework, support both cpu and cuda

Benchmark

use cargo bench to run the benchmark

cargo bench

given a tensor with N elements, and 4 dimensions, the shape is [N/8,2,2,2]

1/6 of the elements are non-zero, the rest are zero.

testing on 4090 GPU and 16 vCPU Intel(R) Xeon(R) Platinum 8352V CPU @ 2.10GHz

N = 2^10

Type	Performance
GPU	40.460 µs
CPU(rayon Parallel)	327.70 µs
CPU(sequential)	2.7161 µs

N = 2^15

Type	Performance
GPU	43.611 µs
CPU(rayon Parallel)	3.5588 ms
CPU(sequential)	66.236 µs

N = 2^20

Type	Performance
GPU	52.509 µs
CPU(rayon Parallel)	78.130 ms
CPU(sequential)	2.2676 ms