Fix RNG segfault related to #297

[jeffdonahue]

This PR removes the caffe_set_rng behavior which was causing a crash (#335) with segfault by constructing a variate generator with a reference to Caffe's RNG engine (something that I found out was possible after hours of googling..).

For some reason, I still get a segfault about 10% of the time I try to run ImageNet training, but this time it happens when filling the biases in conv1 (specifically when SyncedMemory calls malloc to initialize them)...

I0418 10:50:52.465656 16095 net.cpp:75] Creating Layer conv1                                                                                                                                                                        [399/3264]
I0418 10:50:52.465667 16095 net.cpp:85] conv1 <- data
I0418 10:50:52.465687 16095 net.cpp:111] conv1 -> conv1
[New Thread 0x7fffe2db1700 (LWP 16097)]

Program received signal SIGSEGV, Segmentation fault.
0x00007fffef7ee5ae in ?? () from /lib/x86_64-linux-gnu/libc.so.6
(gdb) bt
#0  0x00007fffef7ee5ae in ?? () from /lib/x86_64-linux-gnu/libc.so.6
#1  0x00007fffef7f0f95 in malloc () from /lib/x86_64-linux-gnu/libc.so.6
#2  0x0000000000461c72 in caffe::CaffeMallocHost (ptr=0x9791a0, size=384) at ./include/caffe/syncedmem.hpp:27
#3  0x0000000000461ce7 in caffe::SyncedMemory::to_cpu (this=0x9791a0) at src/caffe/syncedmem.cpp:25
#4  0x0000000000461bb8 in caffe::SyncedMemory::mutable_cpu_data (this=0x9791a0) at src/caffe/syncedmem.cpp:73
#5  0x000000000045f058 in caffe::Blob<float>::mutable_cpu_data (this=0xcfa480) at src/caffe/blob.cpp:67
#6  0x00000000004cc167 in caffe::ConstantFiller<float>::Fill (this=0x936790, blob=0xcfa480)
    at ./include/caffe/filler.hpp:37
#7  0x00000000004d2614 in caffe::ConvolutionLayer<float>::SetUp (this=0x936e60, bottom=..., top=0x8fbb38)
    at src/caffe/layers/conv_layer.cpp:65
#8  0x000000000046419f in caffe::Net<float>::Init (this=0x8fd300, in_param=...) at src/caffe/net.cpp:124
#9  0x0000000000462dc1 in caffe::Net<float>::Net (this=0x8fd300, param_file=...) at src/caffe/net.cpp:31
#10 0x0000000000454c01 in caffe::Solver<float>::Init (this=0x7fffffffda80, param=...) at src/caffe/solver.cpp:39
#11 0x00000000004549e0 in caffe::Solver<float>::Solver (this=0x7fffffffda80, param=...) at src/caffe/solver.cpp:23
#12 0x0000000000415482 in caffe::SGDSolver<float>::SGDSolver (this=0x7fffffffda80, param=...)
    at ./include/caffe/solver.hpp:57
#13 0x0000000000415026 in main (argc=2, argv=0x7fffffffdce8) at tools/train_net.cpp:27

Not exactly sure what to say about this but it seems like it must be unrelated to the RNG since the ConstantFiller does not use it...

[shelhamer]

That we could pass a reference could have been so much better documented by boost. Thanks for hunting this down–along with working, the code is simplified this way.

I'll merge once tests finish running and I do a couple train_net.

[shelhamer]

Although this is only a partial fix, it's better to merge this step in the right direction and then address the newly-revealed bias crash by further debugging and a follow-up PR. This seems to be a stranger crash than we originally thought, but the RNG code is better off with this change anyway.