Use the WRITE_TRUNCATE disposition to upload to BigQuery

[nacnudus]

1. So that the tables are instantly updated with new data, rather than being empty for a few minutes while new data is uploaded.
2. To avoid false alarms when a table is detected as being empty, but is about to be repopulated.
3. To mitigate failures to refresh the data, which is currently catastrophic, because not only is the data not refreshed, but the existing data is wiped.

See the diagram for the current update process, which does a TRUNCATE first, and then a bq load: #736 (comment)

Background to why it was developed in the current way:

• False alarms occur because TRUNCATE and LOAD are in different transactions #491
• Tolerate tables being briefly empty during updates #495

• govuk-knowledge-graph-gcp/terraform/bigquery/check-tables-metadata.sql

  Lines 7 to 12 in 54f9f44

  	-- Raise an alert for tables that have zero rows
  	WHEN row_count = 0
  	-- And that aren't likely to be still being refreshed.
  	-- Annoyingly, it isn't possible to TRUNCATE in the same transaction, so
  	-- tables will briefly be empty until they are repopulated.
  	AND last_modified < TIMESTAMP_ADD(CURRENT_TIMESTAMP(), INTERVAL - 10 MINUTE)

Since #517, there is every reason to use those JSON schemas in the bq load command. Doing so would allow us to use replace=true (which will delete the existing schema), because the schema would immediately be replaced.