Add experiments_column_type "json" for top-level JSON experiments maps

[danielkberry]

Problem

The experiments_column_type = "events_stream" day-0 guard attributes events using each event's own event_extra JSON ($.experiment / $.branch). Those keys are only present on nimbus event rows (enrollment/exposure/etc.) — generic metric events (e.g. contile_impression) carry no $.experiment. So the experiments_column_expr "ignore pre-enrollment day-0 data" guard drops all enrollment-day metric events for events_stream data sources.

For first-run / new-user experiments, where most activity happens on the enrollment day, this badly undercounts exposure and engagement. Observed example: a Fenix sponsored-tiles experiment's control exposure read 16.2% via the events_stream outcome vs 56.4% ground truth (~3.5×), entirely because day-0 impressions were dropped — even though those rows are correctly branch-annotated in the client-level experiments map (event_extra just doesn't carry the experiment/branch on non-nimbus events).

Fix

*.events_stream tables also carry the standard client-level experiments map as a top-level BigQuery JSON column on every row. This adds a new experiments_column_type = "json" that gates day-0 rows on that map:

OR JSON_VALUE(ds.experiments, '$."<slug>".branch') IS NOT NULL

mirroring the "glean" / "native" semantics but for a JSON (rather than repeated-struct) column. No existing type fits: "glean" uses mozfun.map.get_key(ds.ping_info.experiments, ...), which expects a repeated STRUCT under ping_info, whereas events_stream's experiments is a top-level JSON column.

Changes:

• experiments_column_expr and get_sanity_metrics gain a "json" branch
• registered in EXPERIMENT_COLUMN_TYPES, documented on DataSource
• tests for the day-0 guard + sanity metrics; extended the constructor parametrization

All 199 tests pass; ruff lint + format clean.

Follow-up

A companion metric-hub change will switch the events_stream data sources (fenix sponsored_tiles_events_stream, events_stream_awesomebar; desktop protections_popup_events, site_breakage_events) from "events_stream" to "json", and affected experiments will need re-analysis.