Alas, the approach through the functions $g_t$ cannot possibly work. That is seen if one takes e.g. $n=2$, $a_1=4/5$, and $a_2=t=3/5$.
Based on numerical evidence, the following approach promises to work. Without loss of generality (wlog) $n\ge2$ and $a_1\ge\dots\ge a_n\ge0$. Let $a:=a_1$, $b:=a_2$, and $Y:=|S-a_1\eta_1-a_2\eta_2|$. Then $$(1)\quad E Y^2=1-a^2-b^2$$ and $$(2)\quad EY^4=3\Big(\sum_3^n a_i^2\Big)^2-2\sum_3^n a_i^4\ge3(1-a^2-b^2)^2-2(1-a^2-b^2)\times\min(b^2,1-a^2-b^2), $$ since $\max_3^n a_i^2\le\min(b^2,1-a^2-b^2)$. By induction, it is enough to show that $$ E|S_2+Y|^p\le E\Big|\frac{S_2+T_2}{\sqrt2}+Y\Big|^p$$ for all $p\in(2,3)$ and nonnegative r.v.'s $Y$ subject to conditions (1) and (2), where $S_2:= \eta_1 a+\eta_2 b$ and $T_2$ is an independent copy of $S_2$. At that, by well-known results (see e.g. Hoeffding55), wlog the r.v. $Y$ takes at most $3$ values (say $0\le u\le v\le w$ with probabilities $r,s,1-r-s$).
Thus, the problem is reduced to a calculus problem on (say) the minimization of a function of $8$ variables $p,a,b,u,v,w,r,s$ subject to a finite number of restrictions. So, in principle this problem is solvable, but seems very involved computationally.