Fix fragment field-merging validation and reduce allocations

[ghostdogpr]

Summary

• preserve named-fragment identity while collecting fields instead of recursively inlining fragment spreads
• compare nested selections using both return types and the correct overlapping or mutually-exclusive mode
• memoize fragment-name pairs and field-map/fragment pairs with constant-size keys
• stop on the first conflict instead of accumulating error chunks
• add same-object and different-object nested response-shape regression coverage plus an 80-fragment DAG benchmark

Previously, nested fields with the same response name but incompatible leaf types could pass validation beneath mutually-exclusive fragments. Recursing with only the first field type also dropped child fields when the two return object types differed. The validator now resolves and compares both sides independently, while retaining fragment names for O(1) pair memoization.

Benchmarks

Like-for-like JMH runs on JDK 25: 2 forks, 5 one-second warmups, 8 two-second measurement iterations per fork, with GC profiling. Scores show throughput with the JMH 99.9% confidence interval.

Benchmark	series/3.x	This PR	Throughput impact	Allocation impact
80-fragment DAG	396.2 ± 5.8 ops/s	715.0 ± 14.2 ops/s	+80.5% / 1.80x	5.83 → 2.81 MB/op, -51.8%
Existing fragment conflicts	40.67k ± 0.53k ops/s	105.19k ± 2.13k ops/s	+158.7% / 2.59x	68.08 → 38.03 KB/op, -44.1%
FragmentsQueryBenchmark	97.28 ± 1.90 ops/s	106.87 ± 1.98 ops/s	+9.9% / 1.10x	81.39 → 4.26 MB/op, -94.8%

Verification

• sbt fmt
• core test: 623 tests passed
• MiMa passed on Scala 2.12, 2.13, and 3.3
• regression tests cover incompatible nested leaf types beneath both identical and different return object types