bluejay-validator performance optimizations

[swalkinshaw]

Reduce validator allocation overhead — ~28% faster on representative queries

Optimizations:

• Replace HashSet/HashMap with Vec + linear scan in hot paths where N is small (parent_fragments, cycle detection, required arguments, argument equivalence, type overlap checks)
• Eliminate Path Vec allocation by making Path Copy
• Optimize duplicates() to skip BTreeMap allocation when no duplicates found (common case), and avoid intermediate (K,T) vec
• Reuse Cache's fragment_definitions HashMap in FragmentSpreadTargetDefined instead of building a separate HashSet
• Collect errors via &mut Vec in FieldSelectionMerging instead of returning intermediate Vecs

Also adds a criterion benchmark suite for the validation pipeline.

Note: this is a manually curated (with the help of Claude) and cleaned up version of an /autoresearch run