Support for non multiple-of-128 channels

[daquexian]

Related issue: #23

The length of SIMD registers in arm devices is 128-bit, and an element is only 1-bit in BNNs, so only the case that the number of channels is 64 or a multiple of 128 is currently supported. It covers some popular networks, like ResNet, while other networks, unfortunately, cannot be deployed on dabnn.

A possible solution is to unify these unsupported cases into BGEMM:

1. Pad on the channel dimension of weight when reading the model so that the channel of weight is a multiple of 128.
2. When generating the image mat by im2col, pad the mat so that the number of rows is a multiple of 128.
3. Since there is no way to perform zero-pad in BNNs, the padded elements will change the result of BGEMM. Correct the result in the following affine layer.

[daquexian]

Fixed in #49 and #51