Unify Random-Access API and Streaming API into Series::snapshots()

[franzpoeschel]

So far, there were two distinct APIs for accessing Iterations:

1. random-access via Series::iterations
2. step-based access via Series::writeIterations() and Series::readIterations()

This PR introduces a new API Series::snapshots() that returns a unified container API. The unified container API can be implemented internally via random-access and via synchronous access, i.e. by using steps.

This is part of an effort to bring both workflows and their functionalities closer together. While this redesigned container API does not yet go all the way along this road, it lays the fundament for such extensibility by (1) cleanly separating the API from its implementation and by (2) redesigning and cleaning up the implementation of the step-based workflow.

Preliminary/experimental support for reopening closed Iterations is now available, especially for file-based workflows.
Note that there are still caveats about reopening Iterations. Especially in ADIOS2, a step cannot be modified once closed, so writing new data is only possible into a new step. However, the Iteration handle currently does not delete old data definitions upon closing, meaning that there will be invalid data definitions in the handle.
In future, a closed Iteration will be removed from internal data structures (but users may keep it around for longer than that in their own data structures).

Future features to be built on top of this include:

1. Random-access for variable-based encoding based on SetStepSelection() in ADIOS2
2. Going backwards through the steps of an ADIOS2 file (by closing and reopening)
3. Dealing better with duplicate Iterations, arising from checkpoint-restart workflows (e.g. pre-scanning and eliminating duplicates beforehand)
4. Resource management, such as deleting Iteration handles upon closing them, and recreating them upon reopening

writeIterations() and readIterations() are now implemented as aliases/wrappers to Series::snapshots(). Due to the past somewhat unfortunate decision to introduce IndexedIteration as iteration type for readIterations(), the method name cannot be reused without breaking compatibility. writeIterations() however is just a wrapper for snapshots() with default arguments.

TODO:

• Merge Filename extensions: Allow specifying wildcards #1584 first
• Better logic for READ_WRITE mode
• backwards compatibility: using definitions, empty headers
• documentation, change examples
• maybe snapshots(Snapshots::RandomAccess) vs snapshots(Snapshots::CollectiveAccess) or so

• Follow-up: Container API: How to deal with erased Iterations? Answer: Iterations currently only being closed, not erased. Problem for later.
• follow-up: snapshots().begin() -> reset to start, add sth like snapshots().current(), alternatively keep the current semantics and add snapshots().reset()?
• Follow-up: Flush open iterations instead of dirty iterations.

• default: defer_iteration_parsing in READ_LINEAR mode

• follow-up: Check that the last iteration is closed implicitly upon destruction
• follow-up: delete closed iterations in Append mode
• follow-up: re-work access modes, create-linear, append-linear, ..?
• follow-up: Python bindings

Diff: franzpoeschel/openPMD-api@shared-attributable-data...topic-iterator

[franzpoeschel]

The changes introduced with this PR are mostly (1) internal and (2) a new API. I will go ahead and merge this now, as many of the internal changes are the groundwork for new features laid out in the PR description. The new API should be treated as experimental until release.