Templatize SIMDResultHandler on SL, mark handle() final (#5223)

Summary:
Pull Request resolved: #5223

Follow-up to D104552827 per Matthijs's review feedback. The non-template `SIMDResultHandler` base class had a virtual `handle()` using `simd16uint16_tpl<NONE>` in DD mode, while derived handlers used SL-specific types. This meant derived `handle()` methods HIDED rather than OVERRODE the base, preventing use of `override`/`final`.

This restructures the hierarchy: metadata fields from the old `SIMDResultHandler` are merged into `SIMDResultHandlerToFloat` (stays non-template for the polymorphic `FastScanCodeScanner::handler()` interface), and a new template `SIMDResultHandler<SL>` is inserted between them with a virtual `handle()` whose `simd16uint16` type matches the SL parameter. Concrete handlers now genuinely override this virtual and are marked `final`, signaling devirtualization/inlining intent.

Also removes dead code: `with_SIMDResultHandler()` (zero callers), `simd_result_handlers_accept_virtual` (only used by that function), and the file-scope `simd16uint16` alias (only used by the old base class).

Reviewed By: alibeklfc

Differential Revision: D105611944

fbshipit-source-id: 2faaed448692e38459a9c6253039cfa366512a7e

Author: Michael Norris

faiss/IndexIVFRaBitQFastScan.h

@@ -225,9 +225,6 @@ IVFRaBitQHeapHandler<C, SL>::IVFRaBitQHeapHandler(
     }
 }
 
-// Explicit alias — must match SIMDResultHandler::handle() signature.
-using simd16uint16 = simd16uint16_tpl<SINGLE_SIMD_LEVEL_256>;
-
 template <class C, SIMDLevel SL>
 void IVFRaBitQHeapHandler<C, SL>::handle(
         size_t q,

faiss/IndexRaBitQFastScan.h

@@ -134,7 +134,6 @@ template <
 struct RaBitQHeapHandler
         : simd_result_handlers::ResultHandlerCompare<C, with_id_map, SL> {
     using RHC = simd_result_handlers::ResultHandlerCompare<C, with_id_map, SL>;
-    using RHC::handle;
     using RHC::normalizers;
     static constexpr SIMDLevel SL256 = simd256_level_selector<SL>::value;
     using simd16uint16 = simd16uint16_tpl<SL256>;
@@ -188,7 +187,7 @@ struct RaBitQHeapHandler
         }
     }
 
-    void handle(size_t q, size_t b, simd16uint16 d0, simd16uint16 d1) {
+    void handle(size_t q, size_t b, simd16uint16 d0, simd16uint16 d1) final {
         ALIGNED(32) uint16_t d32tab[32];
         d0.store(d32tab);
         d1.store(d32tab + 16);

faiss/impl/fast_scan/fast_scan.cpp

@@ -353,9 +353,6 @@ int pq4_pack_LUT_qbs_q_map(
     return i0;
 }
 
-// declared in simd_result_handlers.h
-bool simd_result_handlers_accept_virtual = true;
-
 int pq4_qbs_to_nq(int qbs) {
     int i0 = 0;
     int qi = qbs;

faiss/impl/fast_scan/rabitq_result_handler.h

@@ -38,7 +38,6 @@ using rabitq_utils::SignBitFactorsWithError;
 template <class C, SIMDLevel SL = SINGLE_SIMD_LEVEL_256>
 struct IVFRaBitQHeapHandler : ResultHandlerCompare<C, true, SL> {
     using RHC = ResultHandlerCompare<C, true, SL>;
-    using RHC::handle;
     using typename RHC::simd16uint16;
 
     const IndexIVFRaBitQFastScan* index;
@@ -83,7 +82,7 @@ struct IVFRaBitQHeapHandler : ResultHandlerCompare<C, true, SL> {
             const FastScanDistancePostProcessing* ctx = nullptr,
             bool multibit = false);
 
-    void handle(size_t q, size_t b, simd16uint16 d0, simd16uint16 d1);
+    void handle(size_t q, size_t b, simd16uint16 d0, simd16uint16 d1) final;
 
     void set_list_context(size_t list_no, const std::vector<int>& probe_map)
             override;