$\newcommand{\mk}{\mathfrak}$Let $\mk{h}_{2n+1}$ be the $(2n+1)$-dimensional Lie algebra (basis $x_1,\dotsc,x_n,y_1,\dotsc,y_n,z$, nonzero brackets $[x_i,y_i]=z$).

The $n$-dimensional abelian Lie algebra $\mk{a}_n$ (basis $a_1,\dotsc,a_n$) acts on it by $a_i\cdot x_i=x_i$, $a_i\cdot y_i=-y_i$, rest of the action being zero.

Through the quotient map $\mk{h}_{2n+1}\to\mk{a}_{2n}$ one thus gets an action of $\mk{h}_{2n+1}$ on $\mk{h}_{4n+1}$. 

Then define the semidirect product $\mk{h}_{2n+1}\ltimes\mk{h}_{4n+1}$, and identify the central basis elements of the two.

This thus defines a $(6n+1)$-dimensional Lie algebra with basis $x_i,y_i,t_i,u_i,v_i,w_i,z$ ($1\le i\le n$) and nonzero brackets
\begin{gather*}
[x_i,t_i]=t_i,\quad [x_i,v_i]=-v_i,\quad [y_i,u_i]=u_i, \quad[y_i,w_i]=-w_i, \\
[x_i,y_i]=[t_i,v_i]=[u_i,w_i]=z.
\end{gather*}

This is supersolvable. The nilpotent radical has basis $t_i,u_i,v_i,w_i,z$ ($1\le i\le n$), so the quotient is abelian with basis $x_i,y_i$ ($1\le i\le n$). Observe that the isomorphic image of $\mk{h}_{2n+1}$, i.e., with basis $x_i,y_i,z$ ($1\le i\le n$) is a Cartan subalgebra (= nilpotent, equal to its normalizer).

This does not split (for $n\ge 1$). Indeed, suppose by contradiction it does: let $\mk{a}$ be a splitting subalgebra. Then the subspace generated by $\mk{a}$ and $z$ is an abelian $(2n+1)$-dimensional subalgebra and is a Cartan subalgebra as well, unlike the previous one which is abelian. But in a solvable Lie algebra all Cartan subalgebras are conjugate (cf. Bourbaki) under automorphisms of the Lie algebra. This is a contradiction (for $n=1$ it can be directly checked: indeed $[x_1+a,y_1+b]$ is never zero for any $a,b$ in the nilpotent radical).

<hr>

Edit 1. Wondering if the above 7-dimensional example is of minimal dimension: <s>I check that the minimal dimension of an example is 6.</s> Here's <s>the</s> a 6-dimensional example (it depends in a nonzero parameter $\lambda$), without details (which are analogous to the previous one): basis $(x,y,t,u,v,w)$, nonzero brackets

$$[x,t]=\lambda t,\quad [x,u]=u,\quad [x,v]=-v,\quad [u,v]=[x,y]=z.$$

[By the next edit, this is not the smallest example. I think this becomes the smallest example if we require in addition that $[\mk{g},\operatorname{nil}(\mk{g})]=\operatorname{nil}(\mk{g})$, where $\operatorname{nil}(\mk{g})$ means the nilpotent radical.]

<hr>

Edit 2. Looking for a reference I found Bourbaki, Lie I, §5, exercise 6:

Consider the 5-dimensional Lie algebra with basis $(u,v,w,x,y)$ and nonzero brackets
$$[u,v]=w,\quad [u,x]=x,\quad[v,y]=y.$$

Check that the subspace with basis $(w,x,y)$ is the nilpotent radical, and that $\mathfrak{g}$ is not the semidirect product of an abelian subalgebra with a nilpotent ideal (hence the nilpotent radical doesn't split, since the quotient is abelian).