Big rewrite

[SimonHeybrock]

This adds, as scippnexus.v2, the future version of ScippNexus. The trigger for this rewrite were significant performance issues with small files that contain thousands of groups and datasets. However, it also attempts to resolve some complicated logic of the old implementation, which did lead, e.g., to hard to understand stack traces in exceptions.

To use this, just try import scippnexus.v2 as snx. The basic API is the same as before.

Apart from internal design and significant performance improvements (not affecting large datasets), there are a number of changes:

• Fallback to loading as DataGroup is not better at preserving "partial" information, if assembly fails.
• Transformation chains are not executed and not resolved. We will provide post-processing functions for this.
• NX*geometry classes are loaded as DataGroup without processing. We will provide post-processing functions for grouping this by detector_number, etc.
• NXevent_data is not automatically grouped by detector_number. There is a post-processing function for this, but we will need to add some that operate recursively.
• Error-fields are "absorbed" into data fields. Accessing them independently is not possible any more. May change my mind about this again...

For review:

• Start with base.py. class File and its helper functions are mostly (but not completely) unchanged.
• Continue with nxdata.py.
• Look at the rest in any order.
• There is some unused code in nxtransformations.py, which I plan to refactor in a follow-up into a couple of user-facing functions for resolving and executing depends_on-chains.

Bugfixes:

• Fields with shape=(0,) are now correctly loaded as datetime if they should. Previously this always returned a time (offset).
• Fields with shape=(0,) are now correctly loaded as transformation if they should. Previously this always returned the plain value of the transformation.

[SimonHeybrock]

Apart from addressing your comments, I pushed another small change the switches out the cached_property for manual lazy init for the Group._children and Group._nexus properties. This should make it easier to see through the logic. I kept cached_property for the other methods/properties that "just" return something simple that is computed.

[jl-wynen]

I honestly don't see the benefit. Was the logic unclear before?

[SimonHeybrock]

I honestly don't see the benefit. Was the logic unclear before?

In particular the _children felt odd: We write to the return value, but it is not simply returning an actual field of the class instance. Rather, it is returning something that it constructed in the property. We basically relied on cached_property to do this exactly once, and this may be unexpected to the reader.

[SimonHeybrock]

Should be ready for another review @jl-wynen.

[SimonHeybrock]

Tracking next steps in #120.

[SimonHeybrock]

@jl-wynen I fixed another small issue, and pushed some extra horrible code to deal with event-data fields embedded in NXdetector, to somewhat restore the ability to load SNS files. I am not planning further changes in this PR, so this is ready for a "final" look.