replace the final 8xupsample with convtranspose

[windyrobin]

in the scenario of gt_down_sample ==1, the header with 8x upsampling will lost much space info, replacing it with deconv will get a better result,

in my dateset, I get 4% performance boosting , FYI https://github.com/windyrobin/TorchSeg/commit/bd3085e50deb34e3d586ada3de6d0209c5ea4021#diff-516fe8f873dafdd94ef2d87e5e51428dR228

before:

----------------------------     mean_IU  75.694% mean_IU_no_back 74.636% mean_pixel_ACC 96.662%
29 09:30:08 Evaluation Elapsed Time: 11.91s

after:

----------------------------     mean_IU  79.444% mean_IU_no_back 78.544% mean_pixel_ACC 97.154%
30 20:57:20 Evaluation Elapsed Time: 13.99s

[yu-changqian]

Thanks for your work and improvement.