CI: Improve performance of Cortex-M7 dry run

[mkannwischer]

The dry run for slothy.targets.arm_v7m.cortex_m7 takes over an hour in CI and is our current bottleneck which does not seem justified for the value it provides. We should investigate if that can be cut down (e.g., by adding more caching).

[mkannwischer]

@mjk-claude