Given a vector space $V$ equipped with an inner product $\langle \cdot,\cdot \rangle:V\times V \rightarrow \mathbb R$ and an ordered basis $\{\hat e_n\}$, a rotation is defined as a linear transformation $R:V\rightarrow V$ which preserves the inner product (i.e. $\langle R(x),R(y)\rangle = \langle x,y\rangle$) and the orientation of the space. Such transformations are represented as matrices $r$ such that $r^\mathrm T r = \mathbb I$ and $\mathrm{det}(r) = +1$; the set of such matrices is called the special orthogonal group $\mathrm{SO}(n)$.
If you would like to remove the requirement that the transformation preserve orientation, then you are talking about the orthogonal group $\mathrm{O}(n)$ which preserves inner products. This group includes inversions as well as proper rotations.
More concretely, we can define an orthogonal transformation as a map which takes an orthonormal basis $\{\hat e_n\}$ to another orthonormal basis $\{\hat g_n\}$, and a rotation as an orthogonal transformation which also preserves the orientation of the basis (i.e. an orthogonal transformation with determinant $+1$). It can be shown$^\ddagger$ that any orthogonal transformation can be written as a rotation $r$ composed with an odd permutation matrix $i$; in other words, one can get from any orthonormal basis $\{\hat e_n\}$ to any other orthonormal basis $\{g_n\}$ by performing a rotation and then, if necessary, an odd permutation of the result.
A symmetric $n\times n$ matrix $M$ can be diagonalized by an orthogonal transformation which maps an arbitrary basis $\{\hat e_n\}$ to an eigenbasis $\{g_n\}$ of $M$. That is, one finds the transformation $t$ which maps $\hat e_n$ to $\hat g_n$; from there,
$$(t^\mathrm T M t) \hat e_n = (t^\mathrm TM)\hat g_n = t^\mathrm T \lambda_n \hat g_n = \lambda_n t^\mathrm T \hat g_n = \lambda_n \hat e_n$$
as a result, $t^\mathrm T M t$ is diagonal in the basis $\hat e_n$. As argued above, such a transformation is orthogonal (as it's just a map from one ON basis to another); but we can restrict our attention to special orthogonal transformations by noting that if $t$ has determinant $-1$, then we can add an inversion $i$; the transformation $i\circ t$ then also diagonalizes $M$ (exercise for the reader), and has determinant $+1$.
As a result, for any $n\times n$ symmetric matrix $M$ there exists a special orthogonal transformation (i.e. a rotation) which diagonalizes it.
Can we always represent it in terms of an infinitesimal generator as a rotation about an axis $R=e^{\theta K}$, where $K^\mathrm T = -K$?
We can always write a rotation in that form, as per Qmechanic's answer. The proof is not trivial, but for intuition one can note that any rotation can be broken up into $N$ smaller rotations, and as $N\rightarrow \infty$ the rotations become infinitesimally close to the identity transformation $R(\theta/N) \approx \mathbb I + \frac{\theta}{N} K$ for some antisymmetric $K$.
However, note that rotations are generically not about any single axis, in the sense that e.g. a rotation in $d=4$ dimensions may leave two axes fixed (so specifying a single axis and an angle is not sufficient to define the rotation). Instead, one should think of a rotation as occurring in a plane, see e.g. this wiki article.
Another way to understand this is that a rotation can be defined by writing down an antisymmetric matrix $K$ and then letting $R=e^K$; this exponential map is surjective, as mentioned above. The antisymmetric matrices constitute a vector space - the Lie algebra corresponding to the Lie group $\mathrm{SO}(n)$ - which has dimension $n(n-1)/2$. In 3 dimensions, $3(3-1)/2=3$ and so the space of antisymmetric matrices is isomorphic to the underlying vector space itself. This means that we can associate each antisymmetric matrix to a vector in $\mathbb R^3$, whose magnitude defines an angle and whose direction defines an axis of rotation. But clearly this construction is restricted to the very special case $n=3$; in any other dimension, we cannot specify a rotation simply via an axis and an angle, because the former requires a specification of $n(n-1)/2$ parameters and the latter provides $n$ of them.
As a final note, if we move to complex vector spaces with conjugate-linear inner products, then replacing symmetric $\leftrightarrow$ Hermitian and orthogonal $\leftrightarrow$ unitary leads to the immediate generalization from $\mathbb R$ to $\mathbb C$.
$^\ddagger$If you're familiar with wedge products, then this can be reformulated as follows: an orthogonal transformation $t:\hat e_n \mapsto \hat g_n$ is such that
$$\hat g_1 \wedge \ldots \wedge \hat g_n = C \hat e_1 \wedge \ldots \wedge \hat e_n$$
with $C=\pm 1$. $C=+1$ corresponds to an orientation-preserving orthogonal transformation, i.e. a rotation. From this standpoint, it is obvious that any orthogonal transformation can be expressed as a rotation composed with an inversion; if $C=-1$, we can simply add an odd permutation $\hat g_1 \leftrightarrow \hat g_2$ to make $C=+1$.
