There *is* an 'identification', i.e., a way to interpret a torsion-free affine connection on $M$ as a section of $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})$ in such a way that every (smooth) section of $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})$ over $M$ corresponds to a unique torsion-free (smooth) affine connection.  You haven't explicitly *described* such an identification, though, so it's hard to say whether $(**)$ is 'correct'.

Here is one way to do it:  Given a torsion-free connection $\nabla$ on an $n$-manifold $M$, and a point $m\in M$ let $\exp_m^\nabla:T_mM\to M$ be the (locally defined) exponential map, which is a diffeomorphism from an open neighborhood of $0_m\in T_mM$ to an open $m$-neighborhood $U_m\subset M$.  Choose a linear isomorphism $u: T_mM\to\mathbb{R}^n$ and let $x = u^{-1}\circ (\exp_m^\nabla)^{-1}:(U_m,m)\to(\mathbb{R}^n,0)$.  Then the $\mathrm{GL}_n(\mathbb{R})$ equivalence class $[x]^2_{m,0}{\cdot} \mathrm{GL}_n(\mathbb{R})$ is a canonically-determined element $\gamma(\nabla)_m$ in the $m$-fiber of the bundle $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})\to M$.  Thus, $\nabla$ canonically determines a section $\gamma(\nabla)$ of $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})\to M$.  

It's easy to check (in local coordinates) that this mapping from connections to sections of the given bundle has all of the desired properties.  In particular, every smooth section $\sigma$ of $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})\to M$ is of the form $\sigma = \gamma(\nabla)$ for some unique smooth torsion-free affine connection on $M$.

Note that the natural affine structure on $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})\to M$ is described as follows:  Let $[x]^2_{m,0}{\cdot}\mathrm{GL}_n(\mathbb{R})$ and $[y]^2_{m,0}{\cdot}\mathrm{GL}_n(\mathbb{R})$ be two elements in the $m$-fiber and suppose that two local coordinate representatives have been chosen so that $[x]^1_{m,0}=[y]^1_{m,0}$.  Then define 
$$
t\,\bigl([x]^2_{m,0}{\cdot}\mathrm{GL}_n(\mathbb{R})\bigr)+(1{-}t)\,\bigl([y]^2_{m,0}{\cdot}\mathrm{GL}_n(\mathbb{R})\bigr)
= \bigl([t\,x + (1{-}t)\,y]^2_{m,0}{\cdot}\mathrm{GL}_n(\mathbb{R})\bigr).
$$
One can check (again, in local coordinates), that this is a smooth affine action on $\check J^2(M,\mathbb{R}^n_0)/\mathrm{GL}_n(\mathbb{R})\to M$ and that 
$$
\gamma(t\,\nabla_1 + (1{-}t)\,\nabla_2) = t\,\gamma(\nabla_1) + (1{-}t)\,\gamma(\nabla_2).
$$

As far as references go, I don't know for sure, but I would not be at all surprised to find that this precise construction is described somewhere in Charles Ehresmann's original papers on natural jet bundles.