This allows for saving budgets in multiple formats simultaneously, without
having to recalculate noises.

This will speed up some operations in the CI.

This changes the test to look for the upstream master in the user
current config and use that as the test reference by default.  This
seems like a more sensible default in general, and it's easy enough to
fall back to HEAD if need be ('-g HEAD').

This also fixes some things from the last commit:

* fixes the MR CI pipeline
* adds better command line documentation to the test
* fixes command help
* fixes bug in test cache pruing

The assumption should really be that the HEAD of the repo is the
current noise reference.  There is therefore no need to store a
special reference hash for the tests.  This simplifies things quite a
bit, in particular by not requiring submitters to update a reference.

This commit drops the reference and modifies the test to just compare
the current code against HEAD by default.  The option to compare
against an arbitrary ref is preserved.

The CI is updated for these simplifications.  The "test" job is now
simple generation of the budgets.  The merge request approval job is
renamed 'review:noise_change_approval`, and it's logic remains mostly
the same: it compares the current budgets against the head of the MR
target branch, and requires the same MR approval to pass.  The overall
MR approval process therefor doesn't change.

This adds a CI special job that runs only on merge requests that update
the test git ref, indicating that a noise has been changed.  The MR will
show present a report of the changed noises, and will block on MR approval.
Once approval has been given, the check_approval job can be re-run, which
should now pass, allow the MR to proceed.

If the noises change but the reference has not been updated, then the CI
pipeline will fail as usual.

Some cleanup and updates on the CI, including:
* creating a common docker image for all jobs
* expose test failure report to merge requests

also:
* add h5 output tests
* use debian sphinx packages instead of pip

implement as __main__ for `gwinc.ifo`

also add to CI.  closes #48

Drop specific comparison to MATGWINC output and move to comparisons to
cached hdf5 traces.  Comparisons to all cached IFO traces will be done by
default.  A PDF report of all discrepancies for all IFOs will be
generated if requested.  inspiral ranges will be calculated if the
inspiral_range package is available.

This adds trace caches (in git-lfs) for all currently supported IFOs,
effectively snap-shotting the current state of the repo.

gitlab CI also updated to use new tests

This patch provides a new nb sub-module that defines classes for managing
and calculating noise budgets.  It provides the following overridable
classes:

nb.Calibration  A noise calibration
nb.Noise        A noise source
nb.Budget       A budget of noises

The Budget class includes a calc_trace() method that will return a traces
dictionary that includes data and trace plot styling for every noise term
in the budget recursively.

The existing included interferometers are updated to define their budgets
using this new interface, and the plot_noises function is updated to
accept the new traces dictionary.

An HDF5_SCHEMA describes how trace dictionaries are encoded into HDF5 files.
The new io module includes functions for writing traces to HDF5 files, and
for reading traces stored in this format.

The command line interface is updated to handle this new structure.

With this tweak, most noises agree within 1 ppb.  Exceptions are
seismic and substrate TE (which still have open bugs in matgwinc).

Arguments added to gwinc.test for setting a tolerance level, and a
list of noises to skip in the comparison.

The CI pipeline will now fail when tests fail.

and move to python3

Error is:

  jobs:test:cache:key config cannot contain the "/" character

addresses #2