threads.Threads(...) constructor hangs in low-memory conditions

[vadimkantorov]

If one of threads in the pool fails to reach the main loop (https://github.com/torch/threads/blob/master/threads.lua#L70) because of failed memory allocations, the thread pool does not notice it and tries to remove all endcallbacks from the queue in the synchronize(...) call https://github.com/torch/threads/blob/master/threads.lua#L258 which leads to a hang.

As a solution, one could split the code executed on threads in two parts: one is initialization, the other is the main loop; and one would check next if any threads failed to execute the initialization part.