Bind lambda puller to batch hardware specs

[nolan1999]

There is a bit of confusion as to what "Hardware" means at the moment:

• the user can specify the hardware they need, so as to filter the workers
• version_config defines how much hardware a version needs, as specs for the AWS job definition
• the AWS job puller's config also has hardware specs, to pull jobs. This is kind of a duplicate of the ones in version_config...~~

Maybe we should also have default hardware per version.

Note that memory can be overridden https://boto3.amazonaws.com/v1/documentation/api/latest/reference/services/batch/client/submit_job.html when submitting a job.

Things as they are now:

• The user might specify hardware requirements
• The batch jobs (cloud workers) get requirements from the user (if defined) or from the version_config (if not defined)
• Local workers hope to have enough, if the users did not specify anything
• GPU model, architecture, memory are not meaningfully used (string match when pulling)

Should do:

• If the user does not specify any requirements, the defaults from version_config should be used
• Do the lambdas even need the version_config then? Check, but apparently only HW and timeout gotten: HW solved by above, timeout could be passed as well (and specifiable by user, and filterable, and sortable)
• The batch puller pulls based on the maximum they can offer; based on the requirements, they override the job definition defaults with the right instance type, memory, and vcpu (which does not necessarily match cpu_cores....)
• For this, the batch environment needs a mapper instance type => (max cpu cores, gpu model, gpu architecture, gpu memory), and a function that based on the requirements returns the cheapest instance type to use
• Local workers work as they do now (the cluster might need a similar mapping as well)

[nolan1999]

Also, do we want to have multiple types of GPUs running, i.e., multiple job definitions (and pulling) in AWS Batch?
At submit_job, can specify instance to use, ...

[nolan1999]

Will need somehow mapper from gpu_archi to instances...