Patch MPFuture to not use shared memory

[borzunov]

Shmem-related crashes like the one below periodically happen on servers (which causes the swarm downtime), bootstrap peers and the health monitor:

Jan 20 03:01:05 petals-bootstrap2 python[32150]: terminate called after throwing an instance of 'c10::Error'
Jan 20 03:01:05 petals-bootstrap2 python[32150]:   what():  could not unlink the shared memory file /torch_32150_2656295120_383
Jan 20 03:01:05 petals-bootstrap2 python[32150]: Exception raised from close at ../aten/src/ATen/MapAllocator.cpp:514 (most recent call first):
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #0: c10::Error::Error(c10::SourceLocation, std::string) + 0x3e (0x7f3f0aaa156e in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libc10.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::string const&) + 0x5c (0x7f3f0aa6bf18 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libc10.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #2: at::RefcountedMapAllocator::close() + 0xd1 (0x7f3ef0d11141 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libtorch_cpu.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #3: THManagedMapAllocator::close() + 0x4b (0x7f3f0b0773ab in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libshm.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #4: <unknown function> + 0x4433 (0x7f3f0b077433 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libshm.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #5: <unknown function> + 0x4ac468 (0x7f3f09cac468 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libtorch_python.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #6: <unknown function> + 0x40465 (0x7f3f0aa89465 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libc10.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #7: c10::TensorImpl::~TensorImpl() + 0x2ca (0x7f3f0aa827ea in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libc10.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #8: c10::TensorImpl::~TensorImpl() + 0x9 (0x7f3f0aa82939 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libc10.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #9: <unknown function> + 0x6f6268 (0x7f3f09ef6268 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libtorch_python.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #10: THPVariable_subclass_dealloc(_object*) + 0x2a5 (0x7f3f09ef6555 in /home/borzunov/.local/lib/python3.10/site-packages/torch/lib/libtorch_python.so)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: <omitting python frames>
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #23: <unknown function> + 0x90402 (0x7f3f0ba90402 in /lib/x86_64-linux-gnu/libc.so.6)
Jan 20 03:01:05 petals-bootstrap2 python[32150]: frame #24: <unknown function> + 0x11f590 (0x7f3f0bb1f590 in /lib/x86_64-linux-gnu/libc.so.6)

Currently it happens once in 1-3 days even on bootstrap peers. On servers, this happens more often. There's also a hypothesis that this happens when people run fine-tuning.

This PR monkey-patches hivemind not to use it, since we don't need the .cancel() functionality in hivemind. This should solve the issue until we ship a proper hivemind fix.

[borzunov]

[I have rolled back the merge due to unexpected server behavior. Actually, they started to crash even more often - very surprising.]