[common] fix bug for dtype convertion function

[abenmao]

No description provided.

[changqi1]

@abenmao I have some concerns about the kernel perf because there is a conditional judgment in every loop. Would use the following logic to fix this bug.

    constexpr int kStep = 16;
    int blockSize = size / kStep;
    int remainder = size % kStep;

    for (int i = 0; i < blockSize; ++i) {
        __m512 input_vector = _mm512_loadu_ps(src + i * 16);
        __m256i output_vector = cvt_fp32_to_bf16(input_vector);
        _mm256_mask_storeu_epi16(dst + i * 16, 0xffff, output_vector);
    }

    if (remainder != 0) {
        __mmask16 mask = 0xFFFF >> (16 - remainder);
        __m512 input_vector = _mm512_maskz_loadu_ps(mask, src + size - remainder);
        __m256i output_vector = cvt_fp32_to_bf16(input_vector);
        _mm256_mask_storeu_epi16(dst + size - remainder, mask, output_vector);
    }