Add scheduled inverse-residue execution and AVX2 fused d4 inverse codelet

[falseywinchnet]

Motivation

• Reduce call/stack overhead in the inverse residue traversal by reusing the forward residue schedule and exposing a linear switch-based dispatch for small-ISA builds.
• Enable predictable control flow for SSE/NEON builds to gain a near-free scheduling win while preparing the ground for ISA-specific fused inverse kernels on AVX2/AVX-512.

Description

• Add INV_RES_SCHEDULE storage and implement build_inverse_residue_schedule() to reverse FWD_RES_SCHEDULE and append the final inverse FWD_OP_BINOMIAL, and call it from build_forward_schedules() by adding if (!build_inverse_residue_schedule()) return false;.
• Implement scheduled executors run_inv_residue_schedule_f32 and run_inv_residue_schedule that iterate INV_RES_SCHEDULE with prefetch hints and a switch-based dispatch, and use them in inverse_residues_recursive_f32 and inverse_residues_recursive when BRUUN_LEVEL == 1.
• Factor the AVX2 inverse depth-3 leaf into a register-returning core codelet_d3_tw_res_inv_avx2_core, add a fused codelet_d4_inv_avx2 that merges two child leaves and applies the q==8 inverse before stores, and adapt codelet_d3_tw_res_inv to call the new core on AVX2 builds.
• Add inverse_acceleration_notes.txt documenting the design decisions, gating strategy, and follow-up tasks.

Testing

• Ran the repository unit test suite including inverse/forward round-trip correctness tests for double and float across a range of transform sizes, and the tests passed.
• Performed automated build-and-test runs for x86 ISA variants (SSE and AVX2) with BRUUN_LEVEL variants to validate both scheduled and fused-inverse paths, and the builds/tests completed successfully.

---

Codex Task