feat: add explicit copying to ProtoString

[rioyu123]

Summary

This implements the explicit copying operation discussed in #441 for 0.10. It lets a string type keep its borrowing From<&str> conversion while providing owned storage to buffa.

ProtoString currently requires From<&str> for every input lifetime. A string library whose conversion borrows from its input cannot satisfy that bound for its owned representation without changing conversion semantics or introducing overlapping implementations.

This adds ProtoString::copy_from_str(&str) with a default through From<String>, removes the higher-ranked From<&str> bound, and uses the copying operation for borrowed JSON strings in non-optional singular fields and for generated view-to-owned conversions. The remote derive and buffa-smolstr example override it to retain their direct string construction path. Binary decoding still uses from_wire; optional/repeated/map/oneof JSON fields keep using the representation's native serde implementation.

Custom types with an efficient From<&str> should override the new method; otherwise these paths now allocate an intermediate String.

The latest additions from @iainmcgin document the fallback allocation cost and pin the copying route with a CountedStr fixture, so tests detect a conversion that produces the right value but bypasses the override.

Compatibility

• Generic callers that inferred From<&str> from S: ProtoString need S::copy_from_str(value) or an explicit conversion bound.
• json_helpers::proto_string::deserialize now requires ProtoString; hand-written implementations that only supplied the old From conversions need to implement the trait or provide their own deserializer.
• Existing trait implementations inherit the new method. Inline/shared-string implementations should override it to avoid the default's intermediate String allocation. Remote-derived types already forward to the inner type's direct conversion.
• The remote derive still requires a universal From<&str> on its inner type. Types with a borrowing conversion can implement ProtoString by hand; this PR does not add a new derive option.
• Regenerate custom-string view code before using a representation without the old bound. Generated code for the default String representation and wire encoding are unchanged.

Tests

• A lifetime-bearing Cow fixture keeps a borrowing From<&str> conversion and implements ProtoString for its owned variant. It cannot satisfy the original supertrait bound.
• Four integration tests cover copying temporary input, view-to-owned across singular/optional/repeated/map/oneof fields, JSON and null defaults, binary and text round trips, and clearing fields.
• Two CountedStr tests verify that view-to-owned conversion calls copy_from_str for every string field shape, including map keys and values, and that borrowed-string/null JSON deserialization uses it.
• Runtime JSON tests exercise a type without any From<&str> implementation; derive and inline-storage tests exercise the copying override.

Rechecked head d1029b1 on Linux with Rust 1.95.0 and protoc 33.5:

• task lint, task test (2,931 passed, 117 ignored, including doc tests), and task doc passed.
• cargo check --workspace --all-features and buffa no-default-features checks, with and without json, passed.
• The custom-types example built and ran its binary/JSON round-trip assertions successfully.
• Regenerating WKT and bootstrap descriptor types produced no changes.
• Logging regeneration produced the same three-line difference on both this head and unmodified main (c28717b); that unrelated difference is not included.

The protobuf conformance suite, MSRV check, and cross-target checks were not rerun in this follow-up.

Closes #441.

[github-actions]

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

[iainmcgin]

[claude code] @rioyu123 the approach looks right for 0.10. Thank you for the design: a provided copy_from_str, the supertrait removal, and regression fixtures that fail to compile on the old code.

I pushed two commits to your branch:

• Merged origin/main (own signed commit), keeping your copy_from_str changes; view.rs picked up main's per-message preserve_unknown_fields change cleanly.
• The changelog fragment and the copy_from_str docs now warn that a custom string type that relies on an efficient From<&str> but does not override copy_from_str pays an intermediate String allocation, and that the remote derive needs no change.
• # Contract now says copy_from_str(s) must equal From::from(s.to_owned()).
• New CountedStr fixture and tests: a counter in copy_from_str (not in From) pins that generated view-to-owned code (singular, optional, repeated, map key and value, oneof) and JSON deserialization call copy_from_str. Without it every fixture passes even if codegen falls back to From<String>.
• The fragment cites (#442, closes #441), and two comments that described the removed bound are corrected.

[rioyu123]

Thanks for the additions, especially the CountedStr tests. They cover a gap in the original fixtures: checking that the override is actually used, not just that the resulting value is correct.

I rechecked d1029b1 and reran lint, workspace tests (2,931 passed, 117 ignored), strict docs, all-features and no-default-features checks, and the custom-types example on Linux. WKT and bootstrap regeneration were clean; logging regeneration produced the same three-line difference as unmodified main.

I've updated the description and marked this ready for review for 0.10.