[rl] add import torch to provisioner bootstrap to avoid concurrent dlopen …

[shuhuayu]

When Monarch spawns sub-processes and multiple actor threads begin unpickling messages concurrently, they all try to import torch at the same time, causing a race condition in dlopen of torch._C.so. This results in the misleading error torch._C is not a package, even though the import works fine when done sequentially. The fix is to import torch in the Provisioner's bootstrap function, which runs once per sub-process before any threading starts, ensuring torch._C.so is fully loaded before concurrent unpickling begins. It's unclear whether this is a PyTorch bug (concurrent dlopen should be thread-safe) or a Monarch bug (imports during unpickling should be serialized), so we've added a TODO to remove the workaround once the upstream fix lands.

[wwwjn]

@shuhuayu You can go ahead and land this fix first, the RL integration test failing is fixed in #3041

[shuhuayu]

@shuhuayu You can go ahead and land this fix first, the RL integration test failing is fixed in #3041

Sounds good. Tested it locally and passed and i'll merge first.