[ hgemm ] Use aligned memory allocation in transpose / padding gemm

- Using unaligned memory may invoke SIGSEGV

**Self evaluation:**
1. Build test:     [X]Passed [ ]Failed [ ]Skipped
2. Run test:     [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: skykongkong8 <ss.kong@samsung.com>

nntrainer/tensor/hgemm/hgemm.cpp

@@ -96,8 +96,8 @@ void hgemm_noTrans_padding_wrt_K(const __fp16 *A, const __fp16 *B, float *C,
   const unsigned int lda = K;
   const unsigned int ldb = N;
 
-  __fp16 *A8 = new __fp16[M * K8_high];
-  __fp16 *B8 = new __fp16[K8_high * N];
+  __fp16 *A8 = alignedMalloc(M * K8_high);
+  __fp16 *B8 = alignedMalloc(K8_high * N);
 
   float16x8_t ZEROS = vmovq_n_f16(0.F);
 
@@ -1257,7 +1257,7 @@ void hgemm_noTrans_fallback(unsigned int M, unsigned int N, unsigned int K,
 
 void hgemm_transB(const __fp16 *A, const __fp16 *B, float *C, unsigned int M,
                   unsigned int N, unsigned int K, float alpha, float beta) {
-  __fp16 *B_T = new __fp16[K * N];
+  __fp16 *B_T = alignedMalloc(K * N);
 
   transpose_neon<__fp16>(N, K, B, K, B_T, N);
 
@@ -1268,7 +1268,7 @@ void hgemm_transB(const __fp16 *A, const __fp16 *B, float *C, unsigned int M,
 
 void hgemm_transA(const __fp16 *A, const __fp16 *B, float *C, unsigned int M,
                   unsigned int N, unsigned int K, float alpha, float beta) {
-  __fp16 *A_T = new __fp16[M * K];
+  __fp16 *A_T =alignedMalloc(M * K);
 
   transpose_neon<__fp16>(K, M, A, M, A_T, K);
 
@@ -1279,8 +1279,8 @@ void hgemm_transA(const __fp16 *A, const __fp16 *B, float *C, unsigned int M,
 
 void hgemm_transAB(const __fp16 *A, const __fp16 *B, float *C, unsigned int M,
                    unsigned int N, unsigned int K, float alpha, float beta) {
-  __fp16 *A_T = new __fp16[M * K];
-  __fp16 *B_T = new __fp16[K * N];
+  __fp16 *A_T = alignedMalloc(M * K);
+  __fp16 *B_T = alignedMalloc(K * N);
 
   transpose_neon<__fp16>(K, M, A, M, A_T, K);
   transpose_neon<__fp16>(N, K, B, K, B_T, N);