refactor(nodes): invocation registration logic

[psychedelicious]

Summary

• Add separate InvocationRegistry class to handle invocation registration. Previously it was all in BaseInvocation and BaseInvocationOutput. This was a bit awkward as logic for execution was mixed up with registration. The actual logic is the same, but the responsibilities are split up in a more logical way.
• Simplify the invocation typeadapter caching to use @lru_cache(maxsize=1) instead of what effectively was a bespoke version of the same logic.
• Update all tests and consumers to use the new methods.
• Improved type annotations for registration logic.

Related Issues / Discussions

n/a

QA Instructions

App should still work. Custom nodes should still work. Works for me. No issues expected as its pretty straightforward.

Merge Plan

n/a

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[jazzhaiku]

Nice change! The code is clearer with these 2 classes separated.
I've left some suggestions for improvement.

[jazzhaiku]

@psychedelicious
I’d recommend using regular instance methods instead of classmethods.
This allows clients to create and manage their own instances, or you can use the singleton pattern if a global instance is needed.

Same convenience, but more flexible and easier to test.

[jazzhaiku]

@psychedelicious

If the invocation allowlist or denylist is changed, the cache should be cleared

If the allowlist/denylist are passed in as arguments, then the lru_cache will handle this clearing automatically, reducing the burden on clients and avoiding bugs if they forget.

[psychedelicious]

I've left some suggestions for improvement.

Thanks! There's no urgency to this PR - I just remembered I've had this branch in my local clone for a couple weeks and figured I should push it. I'll review the suggestions when I have spare time.