Dashboard: freeze the usage metric to gpu_sm_occupancy

[notoraptor]

In master, dashboard still allowed admin-users to choose efficiency metric, and each endpoint could receive a metric parameter.

This PR removes the choice, and hardcodes gpu_sm_occupancy as efficiency metric in one dashboard constant. Goals:

• Simplify code.
• Harmonize code logic: SM occupancy is also the hardcoded metric used in underusage notifications.
• Since efficiency metric is frozen, some SQL queries may be optimized based on that assumption. For example, we can introduce a table with pre-computed efficiency columns without having to deal with a dynamically-selected efficiency metric.

Changes:

• Drop the metric query param from /dash/metrics/{metric_distribution, metric_comparison, rgu_usage, metric_trend, rgu_by_user, jobs}.
• Remove the admin selector and the selectedMetric() plumbing; the backend injects the name into the template so Python and JS cannot drift.
• Metric comparison keeps its own metric2 picker, minus the usage metric (it is the fixed first axis); a saved state pointing at it is migrated.
• Fix the job-table user column tooltip: it shows the user email.

Breaking: a bookmarked /dash URL carrying metric= now gets the frozen metric instead of a 400. /dash is dashboard-internal (not /v0), so there is no client contract; a test pins the behaviour.