Add Timer class unifying CPU and GPU timer and use it in net_speed_benchmark

[kloudkl]

This resolves the concern about timing CUDA codes in the discussions of #128.

http://devblogs.nvidia.com/parallelforall/how-implement-performance-metrics-cuda-cc/
http://docs.nvidia.com/cuda/cuda-c-best-practices-guide/#performance-metrics

[kloudkl]

@mavenlin, I observed no significant difference among pure CPU timer, CPU timer with cudaDeviceSynchronize, and cudaEvent_t based GPU timer. Are these what you called weird results?

[kloudkl]

Reopen to simplify future benchmark works.

[shelhamer]

Please rebase on the latest dev and we'll merge. Thanks.

[kloudkl]

@shelhamer, it has been rebased and polished with the newly added cpplint.

[shelhamer]

So sorry, but this needs another rebase because of a complicated merge. If it's any consolation, this has prepared the reconciliation of the MKL and non-MKL versions of Caffe and brought in support for DAGs, improved documentation, and a better organization of the project.

We are adopting a new development strategy that will not have a constant need for rebasing. It will be documented shortly, but the bottom line is that we will no longer rewrite the history of dev.

[kloudkl]

I am glad that Caffe is approaching version 1.0. The workflow that I used to rebase on the most recent merge is as follows.

git remote update upstream
git checkout -b Timer upstream/dev
git cherry-pick ...[first commit of this issue]
git cherry-pick ...[second commit of this issue]
...
git push -f origin Timer:cuda_timing

Absolutely clean history.

[kloudkl]

@shelhamer, this utility class is rebased and tested again. Please merge it for those who are interested in benchmarking run time. Thanks!

[shelhamer]

@kloudkl Thanks! We're catching up on PRs now, so hope to merge lots of the new developments soon.