Fix AdaSmooth NaN on zero-initialized parameters - #522

Problem

AdaSmooth computes its efficiency ratio by dividing the running signed movement by the running absolute movement:

c = s.sum().abs_().div_(n.sum())  # e_t

On the first step, a zero-initialized parameter (a bias, or any zero-init layer — very common) has not moved yet: p_diff is zero, so n.sum() is 0 and the update is 0 / 0 = NaN, which permanently corrupts the parameter.

import torch
from pytorch_optimizer import AdaSmooth
p = torch.nn.Parameter(torch.zeros(4))
opt = AdaSmooth([p], lr=1e-2)
opt.zero_grad(); ((p - 1.0) ** 2).sum().backward(); opt.step()
print(p)  # tensor([nan, nan, nan, nan]) on main

The convergence tests only use randomly-initialized parameters, so this was never hit.

Fix

Guard the denominator with eps (n.sum() + eps) so a zero total movement yields a finite ratio; the first-step update is then well-defined and the parameter trains normally. For non-zero parameters eps is negligible, so behavior is unchanged.

Added test_adasmooth_zero_initialized_parameter (NaN on the current code, finite after the fix). ruff clean; the existing adasmooth tests pass.