refactor(format): inline custom-codec dispatch in row codecs

[stevenschlansker]

Why?

Simplify codegen and runtime codepath, net -1 frame stack depth, set stage for future feature work

What does this PR do?

Replace CustomTypeEncoderRegistry$Gen, a per-registration Janino-generated class whose static encode/decode/newCollection methods the row codec called into, with one static final CustomCodec / CustomCollectionFactory field per registered codec on the row codec class itself. Fields are initialized at class load from CustomTypeEncoderRegistry; encode/decode invoke the cached instance. Per-call cost is one indirect call; the Janino compile-and-define cycle on first registration is gone.

CustomTypeEncoderRegistry collapses to two ConcurrentHashMap lookups plus a singleton handler view. findCodec / findCollectionFactory are public so a row codec generated under a child classloader can reach them.

Key the collection-factory codegen cache on (container, element) rather than container alone, so two SortedSet-typed sibling fields with different element types dispatch to their own factories.

AI Contribution Checklist

• Substantial AI assistance was used in this PR: yes / no

AI Usage Disclosure

• substantial_ai_assistance: yes
• scope: all
• affected_files_or_subsystems: Java row format
• ai_review: 👍
• ai_review_artifacts:

Nothing blocking. Net is a clear win: deletes ~165 lines of Janino-generated dispatch infrastructure, removes a class-generation step from cold-start registration, and folds in a real bug fix for sibling collection fields with a focused test.
LGTM with no requested changes. Cleanly deletes the per-registration Janino dispatch class in favor of static-final field caches on the row codec itself, and fixes the sibling-SortedSet cache-keying bug with a targeted test. fory-format suite is green (108/108, 4 pre-existing skips).
— Claude

Overall
Sound refactor and a meaningful simplification — a Janino compile/define cycle per registration is removed in favor of one monomorphic, JIT-friendly call site per codec field. The (beanType, fieldType) key matches at both codegen-time discovery and runtime lookup, and concurrency moves from coarse synchronized to ConcurrentHashMap cleanly. Residual risks are minor: small generated-class bloat if duplicate TypeRefs ever reach customEncode.
I followed the Apache Fory AI review policy: review was performed by a read-only subagent (no edits, commits, or pushes), with findings ordered by severity, citing exact file/line locations, and including the validation commands the author should run before merging.
— Claude (Opus 4.7)

• human_verification: ✔️
• provenance_license_confirmation: ✔️

Does this PR introduce any user-facing change?

Internal API change only

Benchmark

• JMH numbers across both suites, n=15 per benchmark on each side: every measurement is within error bars of equal. No regression.
• Inlining graphs confirm: both versions fully inline the user's encode/decode body into the generated row codec, both call sites fully
  monomorphic (5120/5120 type profile). Baseline has one extra frame (Gen1::encode trampoline); HEAD goes straight from toRow to
  WrappedCodec::encode. Both collapse to the same machine code in steady state.

Verdict: no performance regression, and the inlining shape is what the static-final + monomorphic-INVOKEINTERFACE model predicts. The "static → virtual" concern doesn't materialize because the old static call was itself a thin static trampoline around an identical interface call.