fix(riscv): keep directed FP rounding residuals exact

[carlosqwqqwq]

fix(riscv): keep directed FP rounding residuals exact

Summary

Order the operands by magnitude before the Fast2Sum residual calculation so
the correction remains finite at max-finite top-binade midpoints. Add f32 and
f64 regression coverage for the directed and nearest rounding results.

Validation

• Base: 7f162b19aeaa47825c82fb1c77ca042f535124e3.
• git apply --check on the exact isolated base: passed.
• git diff --check: passed.
• Isolated exact RISC-V harness: cargo test --lib -- --nocapture — 181 passed,
  including both midpoint rounding tests.
• Full RAX 7f source archive with all four prepared patches: cargo test --lib —
  8601 passed, 0 failed, 2 ignored.
• Remote Linux, cross-build, and rustfmt/clippy checks passed. The shared
  build + boot job currently fails before reaching RAX code in
  microkernel/src/arch.rs because this runner's Rust/LLVM rejects existing
  sse/avx/avx2/avx512f target features; the exact-base Microkernel run passed.

Closes #330