Every call was identical as it was based on an existing constant `jsonschema_version_str`

I'm going to do this a lot.
Docstrings can be collapsed in all editors and can benefit from markdown.

Everything here is already private, so using long comments has no benefit

`typeshed` disagrees with `jsonschema`, this is just enforcing what `jsonschema` says is true

Produces the same result, but skips the upfront `deepcopy`.
No longer modifying the copy inplace, new objects are created inside the iterator.

https://github.com/python/typeshed/blob/937270df0c25dc56a02f7199f1943fdb7d47aa9d/stubs/jsonschema/jsonschema/protocols.pyi#L11

Previously, using a version below `4.0.1` would still always check first if there was a property. This would not change between checks.

Defining in this style removes the need for as much documentation, since the version guards are very clear when each branch is used.

First non-failing version.
Have left most of the original code in. Planning to migrate & adapt the comments before removing.

#

vega#3547 (comment)

vega#3547 (comment)

**Remove before review**.

Using for quicker feedback loop, where running mypy and all tests are not beneficial

vega#3547 (comment)

Note to self: Fill out comments re individual changes.

Temporary, will remove before review.
Tried to isolate to a single function so that I can reproduce on main

I renamed this from `_SLOW_BENCHMARKS` but forgot to invert the bool lol

Rather than checking a function/constant, a single set of behaviour is defined **once** - depending on `jsonschema` version.
This makes the remaining functionality much easier to reason with.
Also easier to avoid typing issues

Provided links to `json-schema` reference for more info.
 Also some minor edits to docs

Currently, this will only ever return a single result, based on https://json-schema.org/draft-07/json-schema-release-notes

If a `None` reached `_get_schema_dialect_uri`, it would cause a runtime error.
`None` does not have a `get` method.
Confident this is only a theoretical issue, but is now fixed at the source.

The original name is misleading, as this is a factory function.
That is, a new `Registry` is created on each call.
I think this could be impacting performance. Discarding the registry every time doesn't utilise the immutable properties provided by https://referencing.readthedocs.io/en/stable/api/#referencing.Registry

Now very clear what is being wrapped, and where to find more information

All of this is intended to avoid repeating any work that has already been performed.

Fixes:
```
2024-08-23 16:23:02.509 [info] [Info  - 4:23:02 PM] (15104) Loading pyproject.toml file at c:\Users\danie\Documents\GitHub\altair\pyproject.toml 2024-08-23 16:23:02.511 [info] [Info  - 4:23:02 PM] (15104) No include entries specified; assuming c:\Users\danie\Documents\GitHub\altair 2024-08-23 16:23:04.953 [info] [Info  - 4:23:04 PM] (15104) Found 8234 source files 2024-08-23 16:23:50.675 [info] [Warn  - 4:23:50 PM] (15104) Workspace indexing has hit its upper limit: 5000 files
```

Related vega#3556

Both versions of these functions can be targeted by an IDE.
But hovering over the name (anywhere but the actual definition) displays *only* the final docstring on hover.

This duplicates the docs, and adds a **NOTE** comment to be super clear when viewing the code itself.

There's quite a lot in here, so I've left my notes in `_subclasses` temporarily.
- Removed unused `use_json=False` branch
- Evaluate the hash table **once** and not every time `SchemaBase.from_dict` is called

Try running `hatch test --all` on main vs this.
Locally, each version gets a 1.32-1.53x speedup.
Somehow this also includes `3.11`, despite that being the version "zero-cost" exceptions were introduced.

Thinking that moving the checks the the definition of the class (rather than the instance) could reduce a lot of calls.

Ignoring these *first* for visibility, since I'll be making changes that otherwise have no context

This code path will not raise `jsonschema.ValidationError`

Uses the default parameter of `getattr` instead

Eventually, this will replace the need for checking `None` as frequently

Related vega#3547 (comment)

Purely to demonstrate the differences to status quo.
There would be no loss in functionality, this would simply be some small changes to generated code and the removal of instance-level checks

https://github.com/vega/altair/actions/runs/10566291662/job/29272723977?pr=3547

Locally, did a full replacement of class hierarchy - but saw no observable performance improvement

…[str, Any]`

No functional or user-facing change.
This will be to support using types other than `dict`, which isn't strictly required here

See docstring for notes and plan

Since `v2` this keyword has not been part of any schema.
It has only been a property of what became `FieldOneOfPredicate` https://github.com/vega/schema/blob/ef61166f3f95154465c4b3ebdca88e3c2d25b005/vega-lite/v2.0.0.json#L4060

- Bypass `dct.items()` when there are no properties
-Reuse result of `props` lookup, rather than twice per hit

Very minor improvement

If the current schema doesn't trigger branches, but the test schema(s) do - it may indicate the tests need to be updated to better reflect actual

…_URI`

Maybe revisit lazy imports in the future, but for now this is too unrelated to the PR

See 8ca4266

Moving to discussion threads on the PR

Not sure why I wrote this with `...` originally.

Wasn't able to demonstrate a performance improvement

Used in `Chart.from_dict`