SpanGroup(s)-related optimizations

[shadeMe]

Description

We were needlessly performing a serialization-deserialization round-trip just to copy a string Dict of SpanGroups. This PR replaces that with a straight-forward copy of the dict's constituents. To facilitate this, a new keyword argument was added to SpanGroup.copy() that allows the user to pass a new reference document to which the copy is bound.

Improvements in performance can be seen in this PR.

Types of change

Refactor, enhancement

Checklist

• I confirm that I have the right to submit this contribution under the project's MIT license.
• I ran the tests, and all new and existing tests passed.
• My changes don't require a change to the documentation, or if they do, I've added all required information.

[shadeMe]

@explosion-bot please test_slow

[explosion-bot]

🪁 Successfully triggered build on Buildkite

URL: https://buildkite.com/explosion-ai/spacy-slow-tests/builds/199

[adrianeboyd]

Where is this kind of copying happening? I don't really understand the need for new_doc?

[shadeMe]

In SpanGroups.copy() > Doc.copy() > _copy_examples() > Language.evaluate().

Without the new param, the copied SpanGroup instances will point to their original Doc objects. This was not the case previously as the from_/to_bytes() code path initializes the former differently.

[adrianeboyd]

Ah, I understand. Can you switch this to have the same kwargs format as SpanGroups.copy()?

[adrianeboyd]

I'm trying to think if there's something I've missed in terms of the refcounts, but nothing jumps out at me.

Can you also update span_groups.pyi, the docstrings, and the API docs?