Make nodes extensible and weak-referencable

[AstraLuma]

Add standard dunder attributes to (almost) all instances of __slots__

Using the test suite as a benchmark, this change produced no noticeable slow down.

Closes #81

[AstraLuma]

I did not touch src/graphql/validation/rules/overlapping_fields_can_be_merged.py

[Cito]

I did not touch src/graphql/validation/rules/overlapping_fields_can_be_merged.py

Good, that PairSet class is only for internal usage anyway.

[Cito]

Thanks for sending a PR. Not quite sure if we really should add these to all classes, like Location (attached to every node) and make Token (mainly for internal use in the lexer) extendible.
Would like to get @syrusakbary's opinion regarding this PR.

[Cito]

In order to get a better feeling for the memory penalty, I ran memory_profiler while parsing the big_schema_sdl now, and found that for each of the 3 classes in the ast module, it uses 4% more memory, 12% more in total. I think this is pretty significant already. So I suggest relaxing only the Node class where it makes most sense. What do you think?

[AstraLuma]

Sure, I can do that. My use case is primarily about Node anyway.

[AstraLuma]

How do I do the memory profiling?

[Cito]

What I did was loading the github_schema, then running parse() on it, and observed the memory increase in that line using the memory_profiler. You can also try with other functions. But of course there must be many nodes in the game.

[AstraLuma]

Ok, some variations. Using this script, each is an average of 3 sequential runs

• master: 42.2MB / 17.8MB
• This PR, as-is: 44.6MB / 19.9MB
• Variation 1: 43.1MB / 18.5MB
• Variation 2: 43.9 MB / 19.3MB
• Variation 3: 43.9MB / 19.3MB
• Variation 4: 44.3MB / 19.8MB

Ok, variations:

• Variation 1: Dropped changes to Location and Token
• Variation 2: Dropped changes to Token
• Variation 3: Dropped __dict__ on Location and Token
• Variation 4: Dropped __dict__ on Token

I was surprised by the memory impact, even between master and variation 3 (presence or removal of __weakref__). All that memory is just space from the extra slots--__weakref__ is initialized to None, which is a singleton. (Under the hood in CPython, it's an actual value, PyNone; PyObject* is not normally NULL.)

I guess I'll go with variation 1, then.

[Cito]

Thanks a lot for the detailed verification. I agree variation 1 is the best compromize.

[AstraLuma]

Ok, I think that's all squared away.