fix(arbitrary): add Bytes shims so bytes_fields + generate_arbitrary compiles (#88)

[tejas-dharani]

Description:

Closes #88

Root cause

Since #51 changed Any.value from Vec<u8> to bytes::Bytes in 0.4.0,
buffa-types --features arbitrary fails:

error[E0277]: the trait bound buffa::bytes::Bytes: Arbitrary<'_> is not satisfied

The same failure hits any user crate that combines bytes_fields with
generate_arbitrary = true — the struct-level #[derive(Arbitrary)] cannot
satisfy the bound for bytes::Bytes fields regardless of cardinality.

Fix

Add three helpers to buffa::__private (gated on feature = "arbitrary"):

Helper	Field shape	Generated type
arbitrary_bytes	singular	bytes::Bytes
arbitrary_bytes_opt	proto3 optional	Option<bytes::Bytes>
arbitrary_bytes_vec	repeated	Vec<bytes::Bytes>

Codegen now emits #[cfg_attr(feature = "arbitrary", arbitrary(with = ...))]
on every bytes_fields-typed struct field (selecting the right helper by
shape) and on the inner field of bytes_fields-typed oneof variants. Map
values remain Vec<u8> regardless of bytes_fields and need no shim.

Tests

• any_arbitrary_with_bytes_value in buffa-types/src/any_ext.rs — regression
  pin for Any directly (the reported failure).
• New basic_arbitrary_bytes build block in buffa-test compiles basic.proto
  with use_bytes_type() + generate_arbitrary(true). BytesContexts has all
  four shapes (singular, optional, repeated, oneof variant), so compilation is
  the primary assertion.
• bytes_contexts_arbitrary_all_shapes exercises all four shapes at runtime
  under --features arbitrary, calling slice(..) on each result to assert
  the type is concretely bytes::Bytes.

Checklist

• task lint clean
• task test — 1 500+ tests, 0 failures
• cargo test -p buffa-test --features arbitrary passes
• task gen-wkt-types run; only google.protobuf.any.rs changed

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[tejas-dharani]

I have read the CLA Document and I hereby sign the CLA

[iainmcgin]

[claude code] Approving with a review-feedback commit pushed directly to your branch (maintainerCanModify is on; let me know if you'd rather I revert and have you take it). The fix design is exactly right — option (2) from #88, three __private helpers covering all field shapes, with codegen selecting by shape and skipping map values. The threading of use_bytes through FieldInfo/VariantInfo reuses the existing classification cleanly, and the BytesContexts test message exercising all four shapes is targeted.

What I pushed in b4e0e6db (after merging main to pick up #94/#95/#96):

• buffa-build::Config::generate_arbitrary doc — the constraint that the downstream feature must be named exactly "arbitrary" and forward to buffa/arbitrary was documented on buffa_codegen::CodeGenConfig, but almost everyone calls buffa_build::Config from build.rs. Duplicated the guidance there with a [features] snippet and the symptom (cannot find function 'arbitrary_bytes' in module '__private') so the error is searchable.
• Helper-level equivalence tests in buffa/src/lib.rs — replaced the all-zero-buffer test with three tests that compare each helper's output against the underlying Vec<u8> / Option<Vec<u8>> / Vec<Vec<u8>> impl from an identical Unstructured. An all-zero seed deterministically yields empty vectors, which would let a regression that drops the inner Arbitrary call slip through. The integration test in buffa-test stays as a type-compat smoke test (compilation is the primary assertion there).
• debug_assert!(!v.use_bytes) on the boxed-variant branch in oneof.rs — boxed variants are message/group types, never bytes, so the shim is structurally unreachable. The assert locks the invariant in case is_boxed_variant ever broadens.
• CHANGELOG — noted that singular, optional, and repeated fields plus oneof variants are covered (the original wording said "singular" only, which would lead a user with repeated bytes to think they're not fixed).
• CI — added a cargo check --workspace --all-features step. CI didn't exercise --features arbitrary anywhere, which is exactly how this regression escaped in the first place. check rather than test to keep the time delta small.
• Helper rustdoc — note that oneof variants reuse the singular arbitrary_bytes helper.

Verified locally on the merged head: cargo build -p buffa-types --features arbitrary (the originally-reported failure), cargo test -p buffa-test --features arbitrary (292 pass), cargo test --workspace (1560 pass), cargo clippy --workspace --all-features --all-targets clean, RUSTDOCFLAGS=-D warnings cargo doc --workspace --all-features clean, cargo +1.95 fmt --check clean, task gen-wkt-types no drift.

Thanks for the contribution — this unblocks --all-features and the arbitrary-based benchmarks.