Fix log off by 1 #3847

[R-N]

Fix #3847

Additionally, csv loss log was inconsistent with console output.

• csv: epoch 1-based index, step 1-based index
• output: epoch 0-based index, step 1-based index

I made it to be the same as output (epoch 0-based index).

Also, learn scheduler was weird that it used "<=" sign, so 0.01:10 doesn't mean the first 10 steps, but first 11 steps (applies until step 10 in 0-based steps). I changed this to "<" sign.

Screenshots: (debug prints have been cleaned)

Hypernetwork training

Textual embedding training

CSV log