perf(crypto): stream field-element bytes into hashers and transcript

[Oppen]

Summary

• Eliminates the per-field-element heap allocation in Merkle leaf hashing and Fiat-Shamir transcript appends: AsBytes gains a stream_bytes sink method, overridden zero-alloc for the Goldilocks base field and its degree-3 extension, replacing as_bytes()/to_bytes_be() returning a fresh Vec<u8> per element per hash.
• Adds FieldElementVectorBackend::hash_data_parts + verify_merkle_path_from_hash so row-pair openings (verify_opening_pair, verify_composition_poly_opening, verify_fri_layer_openings) hash two slices directly instead of concatenating them into a throwaway Vec first; verify_composition_poly_opening and hash_data now delegate to their row-pair/parts siblings instead of duplicating the hashing loop.
• Follow-up: the degree-3 extension's stream_bytes was calling the base-field override three times (one per limb), landing as three separate Digest::update calls on the guest instead of one. Disassembly confirmed the dyn FnMut sink itself was fully devirtualized (no indirect-call cost), so the actual waste was call count. Now reuses the existing zero-alloc write_bytes_be override into a stack buffer and sinks once.
• Recursion guest: single-query 89.7M → 71.1M cycles, multi-query 2.21B → 1.78B cycles (handoff's baselines were 89.7M / 2.21B; target was ~1.85–1.95B multi-query, beaten by a further margin after the follow-up).

Test plan

• cargo test --workspace --exclude math-cuda (492 prover + 137 stark tests, all green) after each commit
• make test-ethrex
• cargo test -p lambda-vm-prover --lib test_recursion_execute_1query -- --ignored --nocapture (in-VM verify accepts, guest commits vk_digest ‖ output, byte-identical across every change)
• make test-profile-recursion-single / -multi cycle counts confirmed against baseline after each commit
• Guest ELF disassembly (llvm-objdump) confirming stream_bytes devirtualizes and the extension-field batching removes redundant BlockBuffer::digest_blocks/memcpy calls
• Multiple rounds of independent adversarial review (correctness, performance incl. a dedicated missed-optimization hunt, implementation simplicity) against the diff