Add INSERT_TIME and CONCURRENT_INSERT_TIME marker extraction to
extract_metrics.sh. Add corresponding fields to format_for_action.sh
for dashboard publishing.

Add two new jobs and extend system-x-benchmark:

- insert-benchmark: pg_textsearch single-txn inserts and concurrent
  pgbench inserts across all three datasets, with validation and
  dashboard publishing for 6 metric prefixes

- baseline-benchmark: GIN+tsvector using built-in Postgres (no
  extensions), covering CREATE INDEX on populated table, single-txn
  insert, and concurrent insert for all datasets with 9 metric
  prefixes

- system-x-benchmark: extended with insert-load and concurrent
  pgbench insert steps for all datasets, plus format/publish steps
  for the new insert and concurrent metric variants

Three bugs fixed:
- Section headers for concurrent inserts lacked "Benchmark" keyword,
  making them invisible to the awk section filter in extract_metrics.sh
- Single-dataset runs used lowercase dataset slug in section matching
  (e.g. "cranfield Insert" vs "Cranfield Insert") causing case-sensitive
  mismatch
- Cranfield insert timing captured only one INSERT's time (~15ms) instead
  of total; now uses clock_timestamp() bookends around all 1400 inserts
- Added bm25_spill_index() before index size measurement in insert
  benchmarks so pg_relation_size() reflects actual data (not just the
  memtable metapage)

tp_calculate_idf_sum() was called on every tp_insert(), scanning the
entire term hash table to recompute a sum that was never read. Profiling
showed this consumed 94% of insert time: with 100K+ terms in the
memtable, each row insert triggered a full sequential scan of the
dshash, making insert performance O(docs * terms).

The idf_sum field in TpSharedIndexState and total_terms in TpMemtable
were dead code — written but never read by any query or scoring path.
Remove the function, both fields, and all call sites (insert path,
build finalization, recovery).

The on-disk metapage total_terms field is left in place to preserve
the page layout for existing indexes.