fix(uniffi): link-time optimisation reaches the libraries the release ships

[bahdotsh]

What was wrong

The release profile sets lto = true, but cargo runs link-time optimisation on a library only when every crate type in its list can take it. The FFI crate listed ["cdylib", "staticlib", "lib"], and an rlib cannot, so cargo passed no LTO flag at all. Every native library from the UniFFI migration (2025-11-07) to 0.29.0 was built without it, on every platform, and no build said so.

The rule is in cargo's source (CrateType::can_lto and the comment in compiler/lto.rs: "a cdylib/rlib combination won't allow LTO") and is cargo issue rust-lang/cargo#2301, closed in 2017 with a workaround and no warning.

The change

• The crate's types are now cdylib and staticlib. The release build now passes -C lto.
• tools/capi-fuzz was the only thing linking the crate as Rust. It now compiles the crate's own lib.rs under its own package, with its own build script generating the scaffolding from the same interface definition. The fuzz target is unchanged.
• Two guards: the_ffi_crate_builds_no_rust_library refuses a crate type that turns LTO off; the_fuzzer_compiles_the_ffi_crate_with_its_dependencies holds the fuzzer's dependencies and features to the crate's, since the fuzz build only runs nightly.

Measured

macOS arm64, same commit, fresh target directory each, capi on:

	before	after
liboffline_protocol_uniffi.a	137,949,536	23,393,288
liboffline_protocol_uniffi.dylib	11,534,416	10,932,208
exported C and UniFFI functions	482	482
build time / peak memory	1m46 / 1.3 GB	2m45 / 2.1 GB

The static library is what the iOS framework and the npm package carry, so that is where nearly all of the win lands. The iOS and Android numbers will be read off the first release artifacts built this way.

Verified locally

• cargo fmt, cargo clippy --workspace -- -D warnings, clippy without the C binding, cargo test --workspace --lib, rustdoc under -D warnings.
• The uniffi crate's 263 unit tests all still run: it never had a doctest, so dropping the rlib loses none.
• The C consumer check (shared, static, exchange, quick start) and scripts/check-c-library.sh macos-arm64 against the LTO release dylib.
• The Python suite against the LTO release dylib: 1124 passed, 4 skipped.
• Both guards negative-controlled six ways.
• The fuzzer: every seed runs, and cargo +nightly fuzz run --sanitizer none for 60 s on macOS ran 1.4 million inputs with coverage rising to 2015 edges. Fuzzing on macOS did not link before this change; fuzz.yml and docs/bridges/c.md now say so.

Not covered here

• fuzz.yml has never run. I will dispatch it on this branch.
• The mobile profile (opt-level and panic strategy) is a separate PR: no build has ever used minisize, while six documents say the mobile artifact ships panic = "abort".
• A Rust consumer of the crate from crates.io (0.29.0 has none) can no longer depend on it.