Port sqr_clipped_relu.h to NEON

Passed STC (https://tests.stockfishchess.org/tests/view/69fd70c39392f0c317213b99)

LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 315360 W: 80995 L: 80315 D: 154050
Ptnml(0-2): 573, 33687, 88501, 34325, 594

Data from vondele:

```
==== master ====
1 Nodes/second : 278894157
2 Nodes/second : 279192601
3 Nodes/second : 278279517
Average (over 3):  278788758
==== neonnnn ====
1 Nodes/second : 280614075
2 Nodes/second : 279971502
3 Nodes/second : 280613919
Average (over 3):  280399832
```

closes https://github.com/official-stockfish/Stockfish/pull/6820

No functional change

src/nnue/layers/sqr_clipped_relu.h

@@ -26,6 +26,7 @@
 #include <iosfwd>
 
 #include "../nnue_common.h"
+#include "../simd.h"
 
 namespace Stockfish::Eval::NNUE::Layers {
 
@@ -120,6 +121,25 @@ class SqrClippedReLU {
         }
         constexpr IndexType Start = NumChunks * 16;
 
+#elif defined(USE_NEON)
+        static_assert(WeightScaleBits == 6);
+        constexpr IndexType NumChunks = InputDimensions / 16;
+        const auto          in        = reinterpret_cast<const int32x4_t*>(input);
+        const auto          out       = reinterpret_cast<int8x16_t*>(output);
+        for (IndexType i = 0; i < NumChunks; ++i)
+        {
+            const int16x8_t words0 =
+              vcombine_s16(vqmovn_s32(in[i * 4 + 0]), vqmovn_s32(in[i * 4 + 1]));
+            const int16x8_t words1 =
+              vcombine_s16(vqmovn_s32(in[i * 4 + 2]), vqmovn_s32(in[i * 4 + 3]));
+
+            const int16x8_t r0 = vshrq_n_s16(vqdmulhq_s16(words0, words0), 4);
+            const int16x8_t r1 = vshrq_n_s16(vqdmulhq_s16(words1, words1), 4);
+
+            out[i] = vcombine_s8(vqmovn_s16(r0), vqmovn_s16(r1));
+        }
+        constexpr IndexType Start = NumChunks * 16;
+
 #else
         constexpr IndexType Start = 0;
 #endif