cuDNN R2

[shelhamer]

This is the master edition of #1731. See #1854 for the initial cuDNN R2 compatible branch by @slayton58.

Although cuDNN R2 is still experimental it is in the release candidate stage of development, so this PR will ready Caffe for R2. This should not be made the default until R2 is done or we add version switching for compatibility.

Note that R2 brings OS X support.

Compatibility

• create general tensor, but set 4D tensor
• set conv 2D descriptors
• set pooling 2D descriptors
• replace accumulation flags with alpha, beta scaling
• let pooling have padding
• pick fastest by default (according to cuDNN heuristics)

Next Up

• keep R1 compatibility #if CUDNN_VERSION < 20 since the library isn't itself backward-compatible we're not going to take this on ourselves.
• pick algorithm in Reshape() instead of Forward()
• expose choice of the forward convolutional algorithm / preference
• switch to the N-D interface now that Blobs are N-D arrays (for N not necessarily equals 4) #1486 is done
• 3D convolution