Change H5ContextManager to use temp files and atomic operations

[tskisner]

Change H5ContextManager to use temp files and atomic operations

• Remove '*args' from H5ContextManager. The h5py.File constructor
  does not take positional args, so there is nothing to pass. Explicitly
  parse and handle the mode keyword arg and check that the file
  exists if required.

• Unless mode="r", open a temporary file for modification. Add a
  classmethod to close an hdf5 file handle, optionally moving the temp
  file into place with a rename.

• Implement H5ContextManager tests, including for case of replacing
  file while previous inode is open.

• Update use of H5ContextManager throughout sotodlib, ensuring that
  the classmethod close() is used, which moves the temp file into place.

[tskisner]

The strange MPI unit test hang / deadlock is reproducible locally. It is not related to this PR, but will work on fixing it.

[mhasself]

Ah, thanks for figuring this out! @mmccrackan should give the final approval, but this seems like a huge move in the right direction.

[tskisner]

The unit test for these changes is not working as intended (using multiprocessing inside the unit test). I don't think it impacts the conclusions, but I'm going to push an update that moves the test to an independent script.

[tskisner]

The testing of inode preservation when the file is changed is now in a standalone script whose main is called in the unit tests. You can also observe the failure of the test in master branch by installing master and running the script from this branch:

./tests/test_hdf5_atomic_update.py

and confirming that the code times out trying to access the HDF5 file. Then observe how that same executable works with the code in this branch.

@mmccrackan , thanks for any feedback.