Provide cross-platform seed generation

[eelstork]

Previously (#842), the use of an entropy source was added so that caffe's RNG uses a random seed by default. Although @netheril96 suggested that we avoid boost::random_device, I contend that incurring a dependency on boost random is a small price to pay for cross-platform compatibility.

Additionally, is there a practical reason to preserve the fallback solution via getpid ?

[eelstork]

We shouldn't use std::random_device.

In C++11 std::random_device may be implemented in terms of an implementation-defined pseudo-random number engine if a non-deterministic source (e.g. a hardware device) is not available to the implementation. In this case each std::random_device object may generate the same number sequence.

In constrast, Boost specifies that for those environments where a non-deterministic random number generator is not available, class random_device must not be implemented.

[bhack]

@eelstork If you are interested I've removed all boost dependencies other than for python some months ago at #2537

[eelstork]

@bhack although I am interested (and was aware of #2537), I feel that this (removing boost) may be premature. Additionally, the problem at hand (seed generation) inspires caution: C++ 11 isn't always an ideal replacement for Boost.
What is the main reason to remove Boost? Smaller footprint on handhelds?

[bhack]

The effort started to minimize the dependencies for mobile and embedded (specially for forward pass only of trained nets). I.e. in android boost it is not so much a first grade citizen. See #2619

[bhack]

I repeat again. Are we sure that we are not embracing a rearguard position?

[eelstork]

@bhack currently caffe builds default to C++98. As mentioned before I'm looking into a workaround so that we can be more inclusive.
@shelhamer can you have a look at this?

[bhack]

@eelstork I know. But after 6 months only Nvidia guys commented to the introduction of c++11 and Boost removal in c++. Also the big #2610 rely on c++11. As pointed also Tensorflow already use c++11 features. The problem is ignoring instances and not what kind of direction maintainers want to take.

[eelstork]

Ready for review/merge.

[eelstork]

@ronghanghu can you guys have a look at this?