masked_scatter thrust->cub (#56750)
Summary:
Benchmark:
```python
import torch
import itertools
def run50_sync(f):
for _ in range(50):
f()
torch.cuda.synchronize()
run50_sync(lambda: torch.randperm(1000000, device='cuda'))
def benchmark(M):
a = torch.randn(M, device='cuda')
m = torch.randint(1, (M,), dtype=torch.long, device='cuda').bool()
v = torch.randn(M, device='cuda')
torch.cuda.synchronize()
%timeit run50_sync(lambda:a.masked_scatter_(m, v))
for M in (100, 1000, 100000, 10000000):
print(M)
benchmark(M)
```
Before:
```
100
8.65 ms ± 80.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
1000
8.75 ms ± 72.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
100000
9.27 ms ± 87.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
10000000
33.6 ms ± 358 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
```
After
```
100
8.04 ms ± 37.6 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
1000
8.09 ms ± 38.6 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
100000
8.63 ms ± 76.5 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
10000000
31.9 ms ± 298 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/56750
Reviewed By: ailzhang
Differential Revision: D28547564
Pulled By: ngimel
fbshipit-source-id: 83aeddfaf7023f9f9501c6b1e2faf91e8b6277b1
