Avoid duplicate schema source readers

[iuliiasobolevska]

Summary

• discover filesystem, inline, and dependency-jar schema sources once per CodeGen instance
• open dependency jars once to load both schemas and type mappings, with deterministic closure
• rebuild readers only after syntax failure and include dependency-jar schemas in diagnostics
• preserve the public CodeGen(CodeGenConfig) interface while exposing an internal synthetic test seam

Benchmark

A successful 2,001-file schema workload was run in seven alternating fresh JVMs against master and this branch using the installed CLI distribution. Median wall time improved from 0.55 s to 0.50 s (about 9%). The lifecycle regression test verifies the stronger invariant directly: one filesystem reader open/close per schema on success.

Verification

• ./gradlew :graphql-dgs-codegen-core:test :graphql-dgs-codegen-core:integrationTest :graphql-dgs-codegen-core:lintKotlin
• 173/173 CodeGenTest tests pass; no module failures
• adversarial review found no issues
• public bytecode still exposes CodeGen(CodeGenConfig) and generate() unchanged