Skip conflict expansion when there is nothing to infer

[notatallshaw]

I was taking a look at #20578 and noticed that for the common scenario of having no conflicts, the PR did not improve things, and when there are only a few declared conflicts with nothing to infer, the dedup is consistently slower (although not significant in absolute terms).

We can optimize these scenarios though: when no declared conflict names a group, or nothing uses include-group, no sets can be inferred and the whole step can be skipped.

Note: inferred sets are now appended in insertion order instead of hash order, and lock validation compares conflicts order-sensitively, so a project with inferred sets can relock once. I don't know if this is a backwards compatibility issue or not, but also a rustc-hash change re-rolls the inferred set order. An MRE of this issue:

MRE

[project]
name = "demo"
version = "0.1.0"
requires-python = ">=3.12"

[dependency-groups]
cpu = []
gpu = []
dev = [{ include-group = "cpu" }]
test = [{ include-group = "cpu" }]
docs = [{ include-group = "gpu" }]
ci = [{ include-group = "gpu" }]

[tool.uv]
conflicts = [
    [{ group = "cpu" }, { group = "gpu" }],
]

One declared set, 8 inferred; they come out in a different order, so the first lock after upgrading rewrites the array once.