Ensure that benchmarks continue to compile

[alice-i-cecile]

As noticed in #2673, our benchmark code can silently go stale and fail to compile. Obviously, this is bad and we should fix it.

Running the benchmarks for each PR is too expensive and noisy (in terms of time, compute and developer attention), but verifying that they compile is much cheaper.

@cart's opinion is that this should be a separate step in the CI, rather than handled by integrating the CI into the main Bevy cargo workspace.