Rewriting mnist.py for easier understanding

[Datadote]

• Flattened train+eval loop and cleaned code
• Kept fetch_mnist and TinyConvNet
• 91 -> 69 lines
• Similar performance(same settings as original mnist.py 100 steps, bs=64, lr=0.001)
  ---- (original) mnist.py ----
  Train Loss 0.10 | Train Acc. 0.97
  Test Acc: 0.955800
  real 0m46.187s
  user 0m40.215s
  sys 0m6.840s
  ---- mnist_new.py ----
  Train Loss: 0.1529 | Train Acc: 0.945
  Test Acc: 0.963
  real 0m46.551s
  user 0m40.530s
  sys 0m6.819s