Fix #134: (protobuf) proto3 packed-default and oneof field resolution

[cowtowncoder]

Summary

This started as a fix for #134 (incorrect decoding of repeated fields under proto3 syntax, as reported by @vogt31337), and grew to include two more related bugs found while investigating proto3 compatibility more broadly.

1. Proto3 "packed by default" for repeated scalar/enum fields (#134)

proto3 defaults repeated scalar/enum fields to "packed" wire encoding unless explicitly overridden with packed=false. This library only ever honored an explicit packed option on a field, defaulting to unpacked when none was present -- correct for proto2, but wrong for proto3, where that same absence means packed. As a result, content produced by any real protoc-based proto3 encoder (Go, Python, C++, etc.) that relies on the packed default failed to decode.

The fix threads the parsed .proto file's syntax (proto2 vs proto3) down through the schema-resolution pipeline: NativeProtobufSchema (captured from ProtoFile.syntax()) -> TypeResolver -> ProtobufField, which now computes the default packed value as repeated && isProto3 && type.isPackable() only when no explicit packed option is present.

Backward compatibility:

• Schemas without an explicit syntax = "..."; statement (the common case in this repo's existing tests) get isProto3=false, i.e. unchanged behavior -- verified via ProtoFile.syntax() returning null in that case.
• Explicit [packed=true]/[packed=false] options are unaffected and always take precedence, for both proto2 and proto3.
• Message/string/bytes fields remain non-packable regardless of syntax (FieldType.isPackable()).
• Reading is unaffected either way for old data: ProtobufField.isValidFor() already accepts both packed and unpacked wire types for a repeated field.
• The one intentional behavior change: writing a repeated scalar/enum field under a proto3 schema with no explicit packed option now produces packed (spec-compliant) output instead of unpacked. No existing test relied on the old proto3 write behavior.
• All new methods/constructors are additive; no public signatures changed.

2. oneof fields silently dropped during schema resolution

MessageElement (from com.squareup.protoparser) keeps fields declared inside a oneof block in a separate oneOfs() list from regular fields(). TypeResolver._resolve() only ever iterated fields(), so any field inside a oneof group vanished from the resolved schema -- no error, it just could never be read or written. This affects proto2 oneof too, not just proto3, but oneof is far more idiomatic in typical proto3 schemas.

Fix: merge each oneOf.fields() into the field list resolved by TypeResolver, alongside the message's regular fields. ProtobufField's constructor already treated FieldElement.Label.ONE_OF correctly (falls through to non-required/non-repeated, i.e. optional-like) once given the field -- the bug was purely that these fields never reached that constructor at all.

3. FileDescriptorSet-based schema loading crashed on proto3 descriptor sets

Found while reviewing this PR: FileDescriptorSet.FileDescriptorProto.getSyntax() called ProtoFile.Syntax.valueOf(syntax) directly on the raw descriptor string. Real FileDescriptorProto.syntax values (as written by protoc) are lowercase ("proto2"/"proto3"), but the enum constants are PROTO_2/PROTO_3 -- so valueOf("proto3") threw IllegalArgumentException for any proto3-origin binary descriptor set, meaning the FileDescriptorSet.schemaFor()/schemaForFirstType() loading path (as opposed to parsing .proto text directly) was completely broken for proto3.

Fix: map the string explicitly ("proto3" -> PROTO_3, anything else -> PROTO_2, matching the pre-existing null-handling default) instead of relying on valueOf().

Test plan

• Proto3PackedDefault134Test: proto3 implicit-packed read, proto3 explicit packed=false override, proto2 default (unpacked) unaffected, no-syntax-declaration unaffected, and a proto3 write+read round trip.
• OneofFieldResolutionTest: verifies all fields inside a oneof block (plus a sibling regular field) are present in the resolved schema, and round-trip correctly through write+read.
• FileDescriptorSetProto3Test: builds FileDescriptorProto/DescriptorProto objects directly with syntax = "proto3" / "proto2" and verifies FileDescriptorSet.schemaFor() no longer crashes and produces the correct packed default for each.
• Full protobuf module test suite passes with no regressions after each change.