Add python bindings for writing Frames

[tmadlener]

BEGINRELEASENOTES

• Add a python wrapper around the different available Frame writers.
• Add a put method to the Frame wrapper that allows to add collections to the Frame without having to explicitly use cppyy.gbl.std.move. Fixes Improve python interface when adding collections to a frame #432
• Add test cases that write via python bindings and read via c++.

ENDRELEASENOTES

• Needs Make the SIOFrameWriter call to finish non-mandatory #446
• Also add put_parameters (+ tests)
• rebase this onto Fix the pre-commit workflow #449
• Needs Make building SIO tests depend on ENABLE_SIO #457

[tmadlener]

This became quite a big bigger than I initially anticipated, since there seem to be some quirks around how cppyy selects overloads for the parameters part. Given the improvements in the overall "pythonness" that we now have in EDM4hep, I am no longer sure the user visible import structure we have here is the best, but I would put that into another PR.

[hegner]

@tmadlener - so this PR is complete as is?

[tmadlener]

Technically yes. I have to check why it is failing CI. Also assuming that we address a nicer "import interface" later.

[jmcarcell]

I haven't tried this but after this using collection probably crashes right? It's not trivial to see this if you don't have the C++ background

[jmcarcell]

I just tried it and indeed it crashes, for this PR is fine but there should be something to prevent those...

[tmadlener]

I think it depends on how you want to use the collection. Indexing into it will crash, because it is effectively empty. Adding new things to it might work, but I haven't tried it.

[tmadlener]

After a bit of playing things seem to work until you try to write a collection a second time, i.e.

event = Frame()
hits = ExampleHitCollection()
# add hits
frame.put(hits, "Hits")

len(hits)  # == 0
hits.create()
len(hits)  # == 1

frame.put(hits, "moreHits")  # seg-fault

I will update the documentation that collections should not be used after calling put.

[tmadlener]

updated docstring

[jmcarcell]

I feel like here going out of range may be more of a problem than in C++ (this is not related to this PR but just indexing out of range a std::deque):

event = Frame()
hits = ExampleHitCollection()
hits[0] # seg fault
hits[1] # <cppyy.gbl.MutableExampleHit object at 0x55dcd5d30d50>
hits[1].x() # seg fault
hits.create(0, 1, 2. 3, 4)
hits[0] # <cppyy.gbl.MutableExampleHit object at 0x55dcc4324040>
hits[1] # <cppyy.gbl.MutableExampleHit object at 0x55dcd5d572b0>
hits[2] # seg fault

It's inconsistent since the behavior is undefined, could be a bit hard to debug when you go out of range and it may seem to work but then it doesn't later on. I'm not sure if there are options, not checking in C++ I think is a design choice since there is also .at, but in python no one is going to use that probably

[tmadlener]

Yeah in the current form there are not that many handrails in python. Could probably be fixed by another thin python shim that introduces the range check. That could then also be used to raise an error in the use-after-move case.

Created #447 to keep track of this.

[jmcarcell]

Since the overload for floats will always be double, is it worth it to add a float=True that will select the float overload? Otherwise it's impossible to add a float parameter from python, in the rare case someone needs float and not double...

[tmadlener]

Possible now via an additional as_type argument.

[jmcarcell]

Looks good. The interface can be improved like in EDM4hep but that can be done later.