fix(distance): add SIMD target attributes to lambdas in compare_batch

Explicitly annotate batch_impl lambda functions in distance headers with DEGLIB_TARGET_AVX512 and DEGLIB_TARGET_AVX2.

GCC/Clang do not propagate function target attributes into nested lambda closures. Without target annotations on the lambdas, compiling AVX-512/AVX2 intrinsics inside non-AVX512 translation units failed with inlining option mismatch errors.

Author: Neiko2002

cpp/deglib/include/deglib/distance/evp_inner_product.h

@@ -216,7 +216,7 @@ namespace deglib::distances::evp_ip {
                uint32_t dim = *((uint32_t*)qty_ptr);
                const size_t mask_bytes = dim / 8;
 
-               auto batch_impl = [query, mask_bytes, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, mask_bytes, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX512 {
                    const std::byte* q_ones = query;
                    const std::byte* q_negs = query + mask_bytes;
 
@@ -452,7 +452,7 @@ namespace deglib::distances::evp_ip {
                uint32_t dim = *((uint32_t*)qty_ptr);
                const size_t mask_bytes = dim / 8;
 
-               auto batch_impl = [query, mask_bytes, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, mask_bytes, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX2 {
                    const std::byte* q_ones = query;
                    const std::byte* q_negs = query + mask_bytes;
 

cpp/deglib/include/deglib/distance/fp16_ip.h

@@ -128,7 +128,7 @@ namespace deglib::distances::fp16_ip {
             const uint16_t* query = static_cast<const uint16_t*>(query_ptr);
             const size_t dim = *((const size_t*)qty_ptr);
 
-            auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+            auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX512 {
                 const size_t nc32 = dim / 32;
                 size_t offset = nc32 * 32;
 
@@ -268,7 +268,7 @@ namespace deglib::distances::fp16_ip {
             const uint16_t* query = static_cast<const uint16_t*>(query_ptr);
             const size_t dim = *((const size_t*)qty_ptr);
 
-            auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+            auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX2 {
                 const size_t nc = dim / 16;
                 size_t offset = nc * 16;
 

cpp/deglib/include/deglib/distance/fp32_ip.h

@@ -139,7 +139,7 @@ namespace deglib::distances::fp32_ip {
                const float* query = static_cast<const float*>(query_ptr);
                const size_t dim = *((const size_t*)qty_ptr);
 
-               auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX512 {
                    const size_t nc16 = dim / 16;
                    size_t offset = nc16 * 16;
 
@@ -260,7 +260,7 @@ namespace deglib::distances::fp32_ip {
                const float* query = static_cast<const float*>(query_ptr);
                const size_t dim = *((const size_t*)qty_ptr);
 
-               auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX2 {
                    const size_t nc8 = dim / 8;
                    size_t offset = nc8 * 8;
 

cpp/deglib/include/deglib/distance/fp32_l2.h

@@ -139,7 +139,7 @@ namespace deglib::distances::fp32_l2 {
                const float* query = static_cast<const float*>(query_ptr);
                const size_t dim = *((const size_t*)qty_ptr);
 
-               auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX512 {
                    const size_t nc16 = dim / 16;
                    size_t offset = nc16 * 16;
 
@@ -262,7 +262,7 @@ namespace deglib::distances::fp32_l2 {
                const float* query = static_cast<const float*>(query_ptr);
                const size_t dim = *((const size_t*)qty_ptr);
 
-               auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX2 {
                    const size_t nc8 = dim / 8;
                    size_t offset = nc8 * 8;
 

cpp/deglib/include/deglib/distance/uint8_l2.h

@@ -130,7 +130,7 @@ namespace deglib::distances::uint8_l2 {
                const unsigned char* query = static_cast<const unsigned char*>(query_ptr);
                const size_t dim = *((const size_t*)qty_ptr);
 
-               auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX512 {
                    const size_t nc32 = dim / 32;
                    size_t offset = nc32 * 32;
 
@@ -262,7 +262,7 @@ namespace deglib::distances::uint8_l2 {
                const unsigned char* query = static_cast<const unsigned char*>(query_ptr);
                const size_t dim = *((const size_t*)qty_ptr);
 
-               auto batch_impl = [query, dim](const void* const* db, float* out_dists) {
+               auto batch_impl = [query, dim](const void* const* db, float* out_dists) DEGLIB_TARGET_AVX2 {
                    const size_t nc16 = dim / 16;
                    size_t offset = nc16 * 16;