Add DaliIterator

[TakuyaNarihira]

This PR adds a capability of using NVIDIA's data loading library DALI in nnabla.

The following example creates a DALI iterator from nnabla's data iterator. You can also create your own DALI pipeline fed into DaliIterator.

import numpy as np
import nnabla as nn

from nnabla.ext_utils import get_extension_context
nn.set_default_context(get_extension_context('cudnn', device_id='0'))

# Create a dummy data iterator which yeilds two `numpy.ndarray`s.
from nnabla.utils.data_iterator import data_iterator_simple
batch_size = 32
data_shape = (4, 256, 256)
nn_it = data_iterator_simple(
    lambda i: (np.full(data_shape, 0, dtype=np.float32),
                np.full(data_shape, i, dtype=np.int32)),
    512, batch_size, shuffle=True)

# Create a pipeline from data iterator
from nnabla_ext.cuda.experimental import dali_iterator
p = dali_iterator.DataIteratorPipeline(nn_it, device_id=0)
dali_it = dali_iterator.DaliIterator(p)

DaliIterator can be used as following.

# Create a whatever network
v1 = nn.Variable((batch_size,) + data_shape)
v2 = nn.Variable((batch_size,) + data_shape)
y = 0
for i in range(500):
    y += v1 + v2
    y -= v1 + v2
    
# In a data processing loop
for i in range(5):
    # Set inputs from nnabla data iterator, which is slower
    # because it transfers the data to GPU synchronously.
    # v1.d, v2.d = nn_it.next()
    
    # Set inputs from DALI iterator
    # This automatically transfers array values
    # to GPU asynchronously to computation
    v1.data, v2.data = dali_it.next()
    y.forward(clear_buffer=True)