Skip col_buffer allocation when not needed

[sguada]

In the line of #1118 this PR based on the discussion with @longjon and @shelhamer extends the cases when the col_buffer is not needed and a matrix multiplication is enough to compute the result.

[longjon]

Thanks for PRing this. However, see my comment #1118 (comment); I believe your expression is wrong and breaks convolution in certain cases.

Also, I think we should test this rather more aggressively in terms of the parameter space than the existing conv tests (even if we only do this once, offline) as in any case this will add significant subtlety to the code paths.

[sguada]

@longjon agreed, we need a more complete set of tests. Currently I only need the first two cases 1x1 and Single, for which I have added test cases, so we can keep that two in this PR and leave the other 2 for later, or I add them now and test them offline (I think they a less frequent use case).

(pad_h == 0 && pad_w == 0) && 
  ((kernel_w_ == 1 && kernel_h_ == 1 && stride_h_ == 1 && stride_w_ == 1) ||  // 1x1
  (width == kernel_w && height == kernel_h) ||  // Single
  (stride_w == kernel_w && width % kernel_w == 0 && kernel_h == 1) ||
  (width == kernel_w && stride_h == kernel_h && height % kernel_h == 0))

[longjon]

(See #1118 (comment), and let's continue the discussion here.)

If it's not too onerous to gain confidence in its correctness, we might as well use the general expression now.

Since all the tests are for !skip_col_buff_, perhaps we should compute need_col_buff_ instead?

[sguada]

@longjon could we merge this?

[sguada]

@longjon although we could compute need_col_buff_ I think the logic is easier to follow with skip_col_buff_.
The first two cases are clear and very useful, so if you are not comfortable with the last two cases let's merge the first two.

[shelhamer]

Closing since the dev branch is deprecated. Please send PRs to master.