For rigor, you might try to ask on Math.SE or MO.SE. In this answer, we will give a heuristic derivation via discretization. We will use a slightly different notation to connect to usual time-evolution in QM, but the idea is the same:
$$\begin{align}
U_{\lambda}(t_f,t_i)~=~&T e^{-\frac{i}{\hbar}\int_{t_i}^{t_f}\! dt~ H_{\lambda}(t)}\cr
~=~&\lim_{N\to\infty} T\prod_{n=1}^N e^{h_{\lambda}(n)} \cr
~=~&\lim_{N\to\infty} e^{h_{\lambda}(N)}e^{h_{\lambda}(N-1)} \ldots e^{h_{\lambda}(2)}e^{h_{\lambda}(1)} , \cr
h_{\lambda}(n) ~:=~&-\frac{i}{\hbar}\frac{\Delta t}{N}H_{\lambda}\left(t_i+ (n-\frac{1}{2})\frac{\Delta t}{N}\right), \cr
\Delta t ~:=~&t_f-t_i. \tag{1}\end{align}$$
Here $\lambda$ is an external parameter. Therefore, if we are allowed to change the order of differentiation and limit, we formally get OP's formula
$$\begin{align}
&\frac{dU_{\lambda}(t_f,t_i)}{d\lambda}\cr
~=~&\lim_{N\to\infty}\sum_{n=1}^N e^{h_{\lambda}(N)}\ldots e^{h_{\lambda}(n+1)} \frac{de^{h_{\lambda}(n)}}{d\lambda} e^{h_{\lambda}(n-1)}\ldots e^{h_{\lambda}(1)}\cr
~=~&\lim_{N\to\infty}\sum_{n=1}^N e^{h_{\lambda}(N)}\ldots e^{h_{\lambda}(n+1)} \frac{dh_{\lambda}(n)}{d\lambda} e^{h_{\lambda}(n-1)}\ldots e^{h_{\lambda}(1)}\cr
~=~&-\frac{i}{\hbar}\int_{t_i}^{t_f}\! dt~ U_{\lambda}(t_f,t)\frac{dH_{\lambda}(t)}{d\lambda}U_{\lambda}(t,t_i).\tag{2}\end{align}$$