perf(recursion): use inline Blake2b-256 for transcript

[quangvdao]

Summary

Use Jolt’s Blake2 inline implementation for the Fiat–Shamir transcript in non-host (recursion/guest-verifier) builds to reduce zkVM cycle counts.

Key Changes

• Add a Blake2b-256 hasher to jolt-inlines-blake2 (output-length parameterized to 32 bytes; not truncation).
• Switch Blake2bTranscript to use the inline Blake2b-256 backend when feature = \"host\" is disabled.
• Wire jolt-core minimal feature to include jolt-inlines-blake2, with target-specific dependency features so it compiles on non-RISC-V while emitting the custom instruction on RISC-V.

Testing

• cargo fmt --all
• cargo test -p jolt-inlines-blake2 --features host
• cargo test -p jolt-core
• cargo test -p jolt-core --no-default-features --features minimal
• cargo clippy -p jolt-inlines-blake2 --all-targets --features host
• cargo clippy -p jolt-core --all-targets
• cargo clippy -p jolt-core --no-default-features --features minimal --lib --tests

Review Guidance

• The transcript algorithm is unchanged (still Blake2b-256); only the hashing backend changes for recursion builds.
• Focus on feature gating (host vs minimal) and ensuring equivalence between the inline and reference implementations.