Make the MobilityDuck batch benchmark runner turnkey

[estebanzimanyi]

The batch runner bench_mduck.sh reads a load_mduck.sql loader and per-query qNN.sql files that are absent from the repository, so the MobilityDuck side of the cross-platform batch benchmark cannot run.

This makes the MobilityDuck side turnkey, mirroring the MobilityDB runner:

• bench/load_mduck.sql loads the cross-platform portability export — the same vehicles.csv / trips.csv / query_*.csv the MobilityDB runner consumes — into the schema queries.sql expects: Trips and Vehicles plus the Query* selectors and the trip_h3 prefilter column. SRID is resolved by MEOS parsers, never by hand: the trip via tgeompointFromHexEWKB and the static geometries via ST_GeomFromText (which accepts the SRID-tagged EWKT).
• bench_mduck.sh reads each query from the shared queries.sql, split on its -- BerlinMOD Qn: markers, and reports any query the loaded extension cannot run instead of recording a phantom sub-second timing.

At scale factor 0.005 (632 vehicles / 1620 trips) the data loads and every query the extension supports runs and times, including the h3 prefilter (geoToH3IndexSet + eEq + eIntersects). A query that needs a function a given extension build does not expose (for example geoToH3Cell or the eDwithinPairs set-set join) is reported and excluded, so the harness is transparent about extension coverage rather than silently green.

[estebanzimanyi]

Closing per maintainer direction: this duplicates the canonical per-tool loading path. The batch runner should reuse MobilityDuck's own BerlinMOD loader rather than a separate hand-authored one over the portability-export CSVs.