Any algorithm would need $\Omega(\log n)$ queries.
To see this, define $f(k)$ to be the number of queries needed for deciding whether an element $x$ appears at least $a$ times in a sorted array $A$. We assume that $x$ appears in $A[m],A[m+1],\dots,A[M]$, and that $k\triangleq\min\{a, m-1, n-M\}$.
Notice that in these notations we are looking to bound $f(\lfloor n/2 \rfloor)$.
Claim: $f(k)\ge \log k$.
Proof:
Consider the first query made by the algorithm.
If it was done within radius $k/2$ of the interval (i.e. somewhere in $[m-k/2,\ldots,M+k/2]$), and found the median at the spot, then we still need at least $f(k/2)$ queries to decide the problem.
If it was done outside that radius, consider the case where the queried cell did not contain the median. Once again, this leaves us with at least $f(k/2)$ queries to be made.
Continue with induction and you get the $\Omega(\log n)$ bound.