Fix flickering of trait errors deduplicated into the trait file

[mabar]

Possible fix for phpstan/phpstan#14932

When a constant-condition error inside a trait is deduplicated and reported "directly in the trait", the reported error invisibly carried the file path of one of the classes using the trait — and which class was picked depended on the run type (cold, warm, partial).

Symptoms:

• an error appears on one run and disappears on the next, with zero changes to any file (reproduction in if.alwaysTrue when found in traits isn't reported correctly through context of a user class phpstan#14932 (comment));
• a path-scoped ignoreErrors entry for such an error matches only on the runs that happen to pick that class;
• --generate-baseline records a different path depending on which run generated it (usually dev vs CI, it depends on amount and timing of workers).

With the proposed modifications, the error belongs to the trait file. Ignoring works by targeting the trait file's path, matches identically on every run, and survives adding or removing classes that use the trait. --generate-baseline writes the trait path as well.

The first commit makes the winner deterministic - fixes the run-to-run flicker on its own, but still attributes the error to the alphabetically last class that uses the trait.
The second makes the attribution independent of the set of using classes. BC break for single-use traits, but necessary for multiple uses. Baseline now matches behavior of the CLI output.

[ondrejmirtes]

Do not modify existing test, add a new one instead please.

[mabar]

@ondrejmirtes Changed to be as close to original as possible, but I can't keep it identical because the test asserted the exact behavior this PR changes