Use Rc<HashSet> for parent_fragments to avoid cloning in field selection merging

[davejcameron]

Summary

Replaces HashSet<&'a str> with Rc<HashSet<&'a str>> for the parent_fragments field in FieldContext. This eliminates O(n) HashSet cloning for every field - field contexts now share the same HashSet via reference counting, and only fragment spreads (which need to add entries) create new HashSets.

Benchmark Results

Fields	Before	After	Improvement
1	2.04µs	2.20µs	-8% (Rc overhead)
2	8.99µs	8.76µs	+3%
4	16.7µs	15.9µs	+5%
8	33.7µs	31.2µs	+7%
16	65.8µs	60.4µs	+8%
32	132.4µs	120.7µs	+9%
64	261.6µs	237.0µs	+9%
128	520.4µs	475.2µs	+9%

The 1-field case has slight overhead from Rc allocation, but this is amortized across multiple fields which is the common case in real queries.

Test plan

• All existing tests pass
• Clippy passes
• Benchmarks show improvement for typical query sizes