fix abnormal spike : Init res state to 1.0

[Maru-mee]

Hello.

This Pull Request addresses an instability issue during the initial steps of training.

Problem

The current initialization of exp_avg_res_row and exp_avg_res_col to zero causes an abnormally large spike in res_approx at step 0 (e.g. res_approx.mean().items()=100000) , resulting in excessively large parameter updates.

This large initial value leads to two problems:

1. Model Instability
   Certain layers may sustain damage.
   e.g., resulting in unexpected artifacts or unnatural color shifts in the output.
2. Persistent Abnormality
   If the user utilizes a high beta[2] value (e.g., > 0.99), this initial abnormal value persists, severely hindering the optimizer's ability to return to a normal state.

Changing point

Non-zero initial value is required to prevent this immediate instability. This PR sets the initial value of the res related states to 1.0.

This change ensures that training starts from a neutral, stable state, effectively preventing catastrophic model failure at the start of the run.