without_extras is slow for lerobot

[konstin]

Given this pyproject.toml, we spend ~38% of resolver thread time in without_extras

Hot code paths:

uv/crates/uv-resolver/src/candidate_selector.rs

Lines 181 to 185 in 0d57d29

	// In the branches, we "sort" the preferences by marker-matching through an iterator that
	// first has the matching half and then the mismatching half.
	let preferences_match = preferences
	.get(package_name)
	.filter(|(marker, _index, _version)| env.included_by_marker(marker.pep508()));

uv/crates/uv-pep508/src/marker/tree.rs

Line 1143 in bec8468

MarkerTree(INTERNER.lock().without_extras(self.0))

Profiling command, on ubuntu 24.04:

cargo build --profile profiling
rm -f uv.lock
samply record --rate 20000 target/profiling/uv lock -p 3.12

We should improve the performance for lerobot.

[BurntSushi]

One simplistic idea is to create a PEP 508 marker once and store it on UniversalMarker. Then the pep508 method can just return it instead of creating it.

I think the main challenge there is just making sure it gets updated as appropriate since a UniversalMarker has &mut methods.

This strategy is predicated on the idea that marker mutation is much rarer than calling m.pep508(). I think that's probably true, but I'm not certain.

[BurntSushi]

(It might also be a good time to get rid of the &mut methods on all of the marker types completely and just double down on its Copy impl.)

[charliermarsh]

We can maybe also optimize the caller here? It looks like we call marker.pep508() twice for each package.