investigate transactions in the generate tasks job

Discussed in this thread. Deploys or other issues can cause us to stop task generation partway, where the parser project is already updated and/or some tasks are already created, meaning that we're never able to finish generating the rest of the tasks. 

Since task generation can be partial, adding more flags or splitting the job to indicate different stages of the generation is likely only an improvement (and a messy one at that).

The primary risk to consider with transactions is the TransactionTooLarge limit that we've been hitting on the server waterfall – we need the tasks/variants generated to be smaller than this limit. Other risks to keep an eye on are the performance implications of transactions (i.e. that they stop and retry if another operation attempts to modify a document that's in the transaction).