[agent] Optimize BFV dot product accumulation

[tlepoint]

Motivation

• Reduce single-thread runtime of MulPIR by optimizing the hot BFV inner-product accumulation path used when computing encrypted database responses.
• Replace higher-overhead ndarray-based accumulation with a more cache-friendly contiguous accumulator to lower per-coefficient overhead.

Description

• Replace the multi-dimensional ndarray accumulation in dot_product_scalar with a contiguous Vec<u128> accumulator and explicit indexing to reduce iterator/Array2 overhead and improve cache locality.
• Perform per-part/per-modulus inner fused-multiply-adds into the contiguous buffer using the existing fma helper, then reduce u128 accumulators to u64 via modulus operators and reconstruct Poly parts.
• Add a unit test test_dot_product_scalar_step_by to validate strided/step_by iterator usage and ensure correctness for common MulPIR access patterns.
• Remove an unused import and replace expect calls on slices with proper Result error returns when coefficient memory is not contiguous.

Testing

• Ran cargo +nightly fmt --all and formatting completed successfully.
• Ran cargo test and all unit tests and doc-tests passed (72 + 73 tests in crates; overall test suites succeeded).
• Ran cargo clippy --all-targets -- -D warnings and it completed without warnings.

---

Codex Task