Hi @Gyusik_Hwang,
Welcome to the forum and happy 2020 to you!
If you want to use DADA2, use method 2. That is, you should not pre-join your reads prior to importing. DADA2 operates by building an error model on forward and reverse reads separately, using that error model to resolve errors and then using those corrected reads to join paired-ends. By pre-joining your reads you are interfering with the the error-building process and so would have unreliable results.
If you do need to use FLASH pre-joined fastq files for some reason, you should use Deblur instead in qiime2. Deblur operates on single-end reads only with a static error model and so can handle pre-joined reads.
For V3-V4 reads, I find DADA2 to retain a higher number of reads than Deblur. This is because as read lengths increases Deblur becomes more conservative (see here for a more detailed explanation). With shorter regions like V4 I find the 2 to be very comparable.
Hope this helps!