refactor(output)!: emit table rows as NDJSON records, keyed by json tags

[Ilyes512]

BREAKING CHANGE: -o json table output is one JSON object per line instead of
one array, its keys are snake_case from the row type's json tags instead of the
column headings, and each value keeps its own type. A pipeline reading
jq '.[].Name' becomes jq -r .name.

Table took [][]string, so the pretty renderer's display strings were the only
thing the JSON writer ever saw. A count arrived as "12", a timestamp as
whatever the table decided to render it as, and jq 'select(.count > 5)' could
not work on any of it. The array it emitted could not be parsed until its
closing bracket arrived, so a killed run left nothing readable — the property
NDJSON exists to provide, and which the docs already claimed. And the column
heading doubled as the JSON key, so one string was doing two jobs for two
audiences: prose a reader may reword freely, and what a consumer's filter
matches on.

None of the three could be fixed alone. Giving the two audiences separate
inputs answers the key-naming question by itself: a row type with json tags for
the consumer, Column headers for the reader, and a generic Table deriving both
from the same rows — which is also what now guarantees every row has exactly
one cell per header.

template list gains from the split immediately: it emits the timestamps
themselves rather than the "3 days ago" the table shows, and omits a field it
has no value for rather than the "-" placeholder. template update emits
up_to_date beside the status sentence, so a consumer filters on the fact
rather than on a phrase that may be reworded.

Closes #109

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}