fix start and end point selection for percolation

[mivanit]

for a maze generated via percolation, a start point is selected randomly and then another is selected by selecting another from the same connected component. This sometimes selects an isolated cell, and causes the solving algorithm (and other things) to fail, since the start and end point cannot be the same.

Possible tasks:

• find the largest connected component when selecting the start and end points. The probability of there not being any connections in a pure percolation maze with parameter $p$ is $p^{(n-1)^2}$. For context, this works out to 39% for n=4, p=0.1, so it's still likely to fail
• A flag in MazeDatasetConfig to allow mazes without solutions