Fix fork issue with py-lmdb-1.7.0+ / Fix for pytorch dataloader with mutiple workers

If any python process uses an aselmdb backend and that process forks, its child processes using the same aselmdb/lmdb_env object can cause a segmentation fault. This segmentation fault is unavoidable when trying to train models with pytorch dataloaders when num_workers>0.

Aside from the current segmentation fault this maybe was also causing a memory leak during training with pytorch dataloaders + aselmdb.

LMDB env are not compatible with fork() , http://www.lmdb.tech/doc/ "Use an MDB_env* in the process which opened it, without fork()ing.".

However we have been able to use this with readonly databases until this line was changed and now causes a segmentation fault https://github.com/jnwatson/py-lmdb/blob/master/lmdb/cpython.c#L3223

This PR fixes the issue, by adding a gate keeper property around the lmdb_env. Each time the getter compares pids to make sure this process does in fact own the lmdb_env, if not open it again. This makes it so that any fork() of the python process such as that used in pytorch dataloaders will never share lmdb envs between processes.