Generic conv & maxpool

[milancurcic]

This PR makes conv1d and conv2d accessible under a generic name conv.

Same for maxpool, except that here we have to make stride a required argument due to interface ambiguity when it's optional.

An example CNN with the new API is now:

  net = network([ &
    input(784), &
    reshape(1, 28, 28), &
    conv(filters=8, kernel_width=3, kernel_height=3, activation=relu()), &
    maxpool(pool_width=2, pool_height=2, stride=2), &
    conv(filters=16, kernel_width=3, kernel_height=3, activation=relu()), &
    maxpool(pool_width=2, pool_height=2, stride=2), &
    dense(10, activation=softmax()) &
  ])

or, without keywords:

  net = network([ &
    input(784), &
    reshape(1, 28, 28), &
    conv(8, 3, 3), &
    maxpool(2, 2, 2), &
    conv(16, 3, 3), &
    maxpool(2, 2, 2), &
    dense(10, activation=softmax()) &
  ])

This is mainly a style choice; I can see pros and cons to either approach. I'm curious what you think.

[jvdp1]

LGTM. Here are some suggestions

[ricor07]

LGTM

[ricor07]

Great. Just to give an update, I started writing some code for my pr poolings but i think i'll be busy until June.

[milancurcic]

@ricor07 no problem at all, whenever you're ready.