No, such an hermitian metric does not necessarily split, even if $n=2$.
As a matter of fact, vector bundles are a quite rigid sort of objects, while metrics are pretty smooth and can be tweaked locally to have any prescribed look
and a given small enough open subset.
By partitions of unity, it is easy to define metrics that have a prescribed
form on some open sets: if $(U_\alpha)$ is an open covering, and $h_\alpha$ is a metric on $U_\alpha$, and $(\lambda_\alpha)$ is a partition of unity relative to this open covering, then $\sum h_\alpha\lambda_\alpha$ is a metric on $E$ which coincides with $h_\alpha$
on the set $U_\alpha$ deprived from the $U_\beta$, for $\beta\neq\alpha$.
So you can construct a metric on $\bigoplus L_i$ which is not obviously
diagonal if the metric $h_\alpha$ is not diagonal (and the $U_\beta$, for $\beta\neq\alpha$, do not cover $\Sigma$).
Example: Cover the sphere $\Sigma$ by two relatively compact open sets $U_\alpha$ and $U_\beta$
on which all the $L_j$ are trivial, and take non-diagonal metrics $h_\alpha$
and $h_\beta$ on $E$. Then the glued metric on $E$ is non-diagonal on $U_\alpha\setminus U_\beta$.
However, it might happen that the obtained metric on $E$ is diagonal in some
other presentation of $E$. This is the case if all $L_j$ are trivial, say,
and if you take a constant non-diagonal metric. Just change the frame and your metric is diagonal.
However, in the important case where the $L_j$ are pairwise non-isomorphic, and ordered by decreasing degree, the automorphisms of $E$ preserve the filtration
$0\subset L_1\subset L_1\oplus L_2\subset\dots$, hence are "upper-triangular".
Since a matrix which is both triangular and orthogonal is diagonal,
any metric on $E$ which is not obviously split is non-split. (Thanks to WillSavin's for correcting a mistake in the first version.)
At a higher level, given an exact sequence of vector bundles with hermitian metrics, the theory of Bott-Chern forms allows to measure the deviation of this metric to a split one. See the initial paper of Bott-Chern (Acta Math., 1965) or the paper of Gillet-Soulé (Annals of maths, 1990).