[ramondalmau]

Summary

Follow-up to #70 (units audit D2). FleetRunner converted the whole JSON input before any
per-flight error handling, so one flight with an invalid trajectory timestamp (or any
malformed flight object) stopped the run for every flight.

• neats_json_to_flights gains an opt-in failures list. When given, a flight that cannot
  be converted is left out and recorded as a FlightConversionFailure (input index,
  identifying attrs, error). Without it, behaviour is unchanged (Dataiku recipes and other
  direct callers are not affected).
• A top level that is an object or a string is a whole-file error: TypeError, not one
  fake failed flight (audit D1 stays open: accepting a bare object is not decided).
• FleetRunner._load_trajectories and scripts/run_json.py use it: a failed flight becomes
  a Failed at parsing error record and the other flights go on. If every flight fails, the
  run stops before loading the weather and returns only error records.
• Recorded errors carry no tracebacks, so they do not keep the input in memory.
• The units audit report records the status (status table, D2 and D1 notes).

Evidence that nothing breaks

• Default path identical: 148 NEATS JSON files (85,128 flights, 10,995,586 points),
  exact comparison before/after: 0 differences.
• New path: identical output on every valid file; the one file that crashes today
  (ao_json_sample.json, trajectory stored as a list) now records its 5 flights as
  failures instead of stopping the run.
• Tests: 51 new/updated unit tests (no data needed), including an all-flights-fail run
  with non-existent weather paths and a check that the input is freed. New end-to-end test:
  5 golden flights + 1 bad flight → the 5 golden results unchanged + 1 error record.
• Full suite with BADA and weather:
• Review: 3 independent reviewers + a re-check; all findings fixed or noted below.
• ruff, ruff format, project mypy clean.

Known, not in this PR

• Low-altitude results are recognised by the text "low-altitude" in the error message
  (pre-existing); a structured marker would be more robust.
• Pre-commit mypy hook is broken for everyone (numpy 2.5.4 stubs vs Python 3.11 target);
  commit made with SKIP=mypy, project mypy passes. Separate chore PR to follow.