feat(rvv): add RVV kernels for fvec_add/fvec_sub and PQ dsub=2 distance tables

[ihb2032]

Summary

On RISC-V, fvec_add, fvec_sub and compute_PQ_dis_tables_dsub2 are dispatched through with_simd_level_256bit, whose level mask (AVAILABLE_SIMD_LEVELS_AVX2_NEON) has no RISCV_RVV bit. RVV hosts therefore fall back to the scalar NONE implementation.

This PR adds native RVV kernels for these functions and fixes the dispatch, following the same vector-length-agnostic approach as the RVV kernels already in the tree: the kernels are __riscv_vsetvl-based (m8) and adapt to any VLEN at runtime instead of assuming a fixed vector width.

Changes

• Add RVV kernels in faiss/utils/simd_impl/distances_rvv.cpp:
  • fvec_add (vector + vector and vector + scalar)
  • fvec_sub
  • compute_PQ_dis_tables_dsub2 (PQ distance tables for dsub = 2, both L2 and inner product). Each product/square is rounded before the final add — like the NONE implementation and unlike a fused vfmacc — so the tables are bit-identical to the scalar ones.
• Add AVAILABLE_SIMD_LEVELS_A0_NO_AVX512 mask in faiss/impl/simd_dispatch.h (A0 minus AVX512: functions with NONE/AVX2/ARM_NEON/RISCV_RVV implementations but no 512-bit specialization, so AVX512 machines keep the AVX2 path).
• Route the four dispatchers in faiss/utils/distances_dispatch.h through the new mask.
• Enable the PQ dsub = 2 distance table path in ProductQuantizer::compute_distance_tables / compute_inner_prod_tables on RISC-V (COMPILE_SIMD_RISCV_RVV), previously guarded to AVX2/ARM_NEON only. No behavior change on other platforms.

Performance

Measured on SG2044 (64-core RISC-V, RVV 1.0), Faiss C++ benchmark suite, rcq-search case:

Metric	Baseline	Patched	Change
instructions	7.25 T	6.87 T	-5.2%
cycles	4.60 T	4.30 T	-6.5%
wall time	238.1 s	223.6 s	-6.1%
branches	334.6 G	305.0 G	-8.8%

fvec_add hotspot in perf record drops from 1.06% (scalar NONE) to 0.18% (RVV) of samples.

Correctness

• sq-accuracy reconstruction errors bit-identical to baseline on all 9 quantizer types, ndiff_for_idempotence=0.
• The compute_PQ_dis_tables_dsub2 kernel is bit-identical to the scalar implementation by construction (no fused multiply-add).