Allow testing the scheduled sets

See merge request !5

This makes post-hoc log analysis easier.

SCHEDULED_POLLERS works like SCHEDULED_PROPORTION, but for polling
processes. 0 means no processes will poll for scheduled jobs; 1 means
all will.

SCHEDULED_THRESHOLD was confusing as 1 meant no jobs were scheduled, and
0 meant all jobs were. SCHEDULED_PROPORTION is the inverse, which is
clearer (to me, at least).

SCHEDULED_ENQ can be one of:

- sidekiq_5: the Sidekiq 6 enqueuer.
- atomic: Heinrich's atomic enqueuer.
- Nothing (the default): the Sidekiq 6 enqueuer.

SCHEDULED_ENQ can be one of:

- sidekiq_5: the Sidekiq 6 enqueuer.
- atomic: Heinrich's atomic enqueuer.
- Nothing (the default): the Sidekiq 6 enqueuer.

SCHEDULED_THRESHOLD is a float (as a string) that sets what proportion
of jobs are scheduled. 1 (the default) means no jobs are scheduled, 0
means all jobs are scheduled, 0.5 means half of jobs are scheduled, etc.

Without this, startup with `run_experiment.rb` will fail silently.

Add more Prometheus config to gitlab.rb

See merge request !4

We can set the scrape interval in Omnibus, and we can also add a scrape
config for process-exporter.

Ping redis to see job rates

See merge request !2

Make it easier to switch scenarios

See merge request !3

Various small improvements

See merge request !1

Making it a bit more repeatable.

This adds a script that runs the entire experiment

Use non-strict ordering, to force randomization of queues like in gitlab, which is also necessary for testing sampling

The p10 is a bit of a fudge, but the rest are broadly right.