fix(ffi): dispatch P-256 signing to the matching ECDSA curve

[canesin]

P-256 ECDSA shares failed in tss_sign_new with ProtocolMismatch because all DKLs23 sessions used the secp256k1 constructor. Dispatch by ciphersuite so P-256 uses new_dkls_r1 and secp256k1 retains new_dkls.

This staging branch merges #5 and adds separate follow-up commits for formatting, C ABI regression tests, the benchmark dependency audit fix, and reproducible fuzz dependency resolution. Both ECDSA curves exercise two-party DKG, signing, signature verification, changed-message rejection, and key-handle ownership.

Validation:

• RED: the new P-256 test fails on the main-branch implementation with ProtocolMismatch; secp256k1 passes.
• Formatting and independent full-diff review pass.
• All 192 workspace tests pass in debug and release, including both new curve regressions. Clippy with warnings denied and benchmark smoke tests pass. Affected FFI release tests and Clippy were rerun after the lockfile update.
• Independent review found no actionable defects in the full diff or the dependency/workflow follow-ups.
• The standalone fuzz workspace previously resolved incompatible upstream FROST dependencies. CI now seeds it from the tested root lockfile. All three local 60-second fuzz smoke runs pass (deserialization, signing, and handles); final-head CI passes build/tests, audit, all fuzz targets, benchmarks, generated headers, ARM64, Android, and CodeQL. Only non-required coverage remains in progress at merge time.
• Updated only the benchmark dependency crossbeam-epoch from 0.9.18 to 0.9.20 to resolve RUSTSEC-2026-0204. cargo audit now reports zero vulnerabilities; existing unmaintained/unsound/yanked dependency warnings remain.

Closes #4. Supersedes #5.