manager_integ_tests: added LocalSGD integration test

[d4l3k]

This adds an integration test for LocalSGD that tests one of the replicas crashing and recovering.

This runs for 8 steps with sync_every=2 and 4 manager steps.

This mostly is refactoring the integration tests into a more flexible Runner class and adds a new local_sgd_train_loop. It does fix one bug in LocalSGD where we don't detach the gradient before running the all reduce.

cc @Jackmin801

Test plan:

pytest torchft/manager_integ_test.py -k local_sgd -s -x