[nicolaskagami]

Fixes: #10855

Context

Currently, two nexuses can race to write to the bootstore and an older configuration of RackNetworkConfig can be written atop a new one. This can be solved with a minimal tweak of how we use the bootstore generation number. If we guarantee that:

• Nexus observes the gen number before it observes the networking config input.
• Only one nexus can write having observed a specific gen number.

Then a newer write must have observed a newer config.

What does this PR change?

• Read the current bootstore gen number in the same transaction as RackNetworkConfigInput.
• Makes bumping the bootstore gen number a "compare-and-swap" approach, so only one instance can successfully do it for a given gen number.
• Only write to the bootstore if we successfully bumped the gen number.

Details

• We can't know if the nexus that lost the race actually had a newer configuration. Currently it will try again next time, but we could prefer to retry immediately.
• I've tested this on voxel making the race much more likely by adding a random 0-3s delay between reading and writing. With that setup, main wrote older configs 19/78 times while this branch never did.