fix: add tail-case handling for elementwise_add_f16x8_pack_kernel to …

[lhycms]

Background

In elementwise_add_f16x8_pack_kernel, each thread processes 8 half elements at once. However, when the input length N is not divisible by 8, the last thread may perform out-of-bounds memory access.

Changes

Added a tail-case handling branch to safely compute the remaining elements one by one:

} else {
    for (int i = 0; nx + i < N; ++i) {
        d_c[nx + i] = __hadd(d_a[nx + i], d_b[nx + i]);
    }
}

Impact

1. Prevents potential out-of-bounds access when N % 8 != 0.
2. Keeps vectorized performance intact when N is a multiple of 8.

[DefTruth]

LGTM