[ML] Adjust model memory limit for module jobs

[pheyos]

Currently, some jobs created via ML modules have a too high model memory limit configured (e.g. 256mb for metricbeat jobs). This could cause issues on low memory clusters (e.g. cloud with a free 1gb ML node) during module execution where all jobs are created and then started almost at the same time.

Suggestion: Revisit the configured model memory limit.

cc: @sophiec20 @zanbel

[elasticmachine]

Pinging @elastic/ml-ui

[sophiec20]

Multi-metric wizard uses following estimation for partition field detectors with metric functions -- this is an approximation, so also round up:
10mb + (64k * num partitions * num detectors) + (10k * num influencers)

If we assume 100 hosts as a high upper water mark and 10 event.datasets then for the following metricbeat jobs (fieldnames are not perfect ECS):

• max disk utilization
  • max(filesystem.used.pct) partition=host.name inf=host.name
  • 25mb
• metricbeat outages
  • low_count partition=event.dataset inf=event.dataset
  • 15mb
• high cpu iowait
  • high_mean(cpu.iowait.pct) partition=host.name inf=host.name
  • 25mb