I think we have $k = q$ (for all $r$).
First, suppose $r = n - 1$. The vector space $V$ can be covered by $q + 1$ codimension-one hyperplanes $W_1,W_2, \ldots,W_{q+1}$ (but no fewer). (If we think of the projective space $[V]$, the most efficient covering is by taking the hyperplane at infinity, plus a set of $q$ parallel hyperplanes in affine space.) Then every $1$-dimensional subspace intersects (in fact, is contained in) some $W_i$. So $k = q$ when $r = n-1$.
To prove $k \le q$ in general, fix a subspace $Z$ of $V$ that has dimension $r + 1$, and use the previous case to choose $q + 1$ subspaces $W_1,\ldots,W_{q+1}$ of dimension $r$ in $Z$, such that every $1$-dimensional subspace of $Z$ is contained in some $W_i$. Since every subspace $U$ of dimension $n - r$ in $V$ contains a subspace of $Z$ that has dimension (at least) $1$, we conclude that $U$ intersects some $W_i$. So $k < q + 1$, as desired.
Now we prove the matching lower bound: $k \ge q$. Let $W_1,W_2, \ldots,W_q$ be any $r$-dimensional subspaces of $V$. Since (as mentioned above), it takes $q + 1$ hyperplanes to cover all of $V$, we know there is some $1$-dimensional subspace $L$ that is not contained in $\bigcup_i W_i$. Let $\overline{V} = V/L$. By induction on $n - r$, there is an $(n-r-1)$-dimensional subspace $\overline{U}$ of $\overline{V}$ that has trivial intersection with every $\overline{W_i}$. Lift $\overline{U}$ to an $(n-r)$-dimensional subspace of $V$. Then $U$ has trivial intersection with every $W_i$.