Create separate implementation for 1x1 convolutional layers.

[Clemens-Alexander Brust]

1x1 convolutional layers can use a much more simple implementation than arbitrary size convolutional layers. This may also improve performance.