Make contiguous case for chainerx::AddAt faster

[emcastillo]

Reworked it similar to #8295

IndexableArrays uses the 1D specialized template whenever it is possible.

For 40 runs of a pathological case, execution time decreased from 0.18 seconds to 0.02

Please note that addat is not currently exposed to the python API.

def bench(name, ary_fn):
    np.random.seed(42)
    x = np.zeros((20, 96, 132)).astype(np.float32)
    y = np.random.rand(20, 96, 36, 11).astype(np.float32)
    indices = np.random.randint(132, size=(36, 11)).astype(np.int32)
    a = chx.array(x, device='cuda:0')
    b = chx.array(y, device='cuda:0')
    indices = chx.array(indices, device='cuda:0')

    for i in range(10):
        cuda = chx.addat(a,indices,b,axis=2)

    cupy.cuda.device.Device().synchronize()
    start = time.time()
    for i in range(40):
        cuda = chx.addat(a,indices,b,axis=2)
    cupy.cuda.device.Device().synchronize()
    take = time.time() - start
    print(name, take)

[emcastillo]

PTAL 😄

[asi1024]

Jenkins and flexCI, test this please.

[chainer-ci]

Jenkins CI test (for commit d85cc1d, target branch master) failed with status FAILURE.

[asi1024]

Jenkins and flexCI, test this please.

[chainer-ci]

Jenkins CI test (for commit d85cc1d, target branch master) succeeded!

[asi1024]

LGTM.