[ramondalmau]

Problem

A FleetRunner does not build its steps (parser, BADA performance model, emissions, climate)
once at the start. Each worker process builds them the first time it needs them and keeps
them in a per-process cache, so the next batch of flights reuses them.

That cache was looked up by step name only ("neats" parser, "bada" performance, ...), not by
the settings the step was built with. So if a second runner ran in the same Python process,
it got the steps built for the first runner, with the first runner's settings.

Example (both runners in one Python session, njobs=1):

r1 = FleetRunnerLargeEmitter(FleetRunnerParams(..., njobs=1))                       # no airport fuel file
r1.eval()
r2 = FleetRunnerLargeEmitter(FleetRunnerParams(..., airport_fuel_path=AF, njobs=1))
r2.eval()

Before this fix, r2 used r1's parser, built without the airport fuel file, so AF was silently
ignored. In the opposite order, the second runner applied a file it was not given. The same
holds for any step setting (bada_path, emission or climate parameters).

Who is affected

Only when one Python process runs two or more runners with different settings, AND the steps
run inside that process or in reused workers:

• njobs=1;
• prefer="threads";
• shutdown_executor_between_steps=False;
• njobs=-1 on a machine or job with only one usable CPU (joblib then runs everything in the
  calling process).

Typical cases: a notebook, a script comparing two configurations, the test suite.
A normal production run (one runner per process, default settings, several CPUs) was not
affected, so past results are unchanged.

Fix

Each runner's step gets its own token when the runner is created. The cache keeps the token
next to the step: a runner with another token builds its own step. Within one runner the
token is the same in every worker, so each worker still builds each step only once
(no slowdown). At most one step per type is kept in memory, as before.

Checks

• 4 new tests (tests/test_step_cache.py), including two real FleetRunners in one process
  with and without the airport fuel file, in both orders. The 3 bug tests fail on the old code.
• Full test suite on this branch (rebased on dev with fix(fuel): reject a q_fuel outside 35-50 MJ/kg (unit error guard) #77): 325 passed, 0 failed.
• Independent review: no defects.
• Audit: new item I11 in docs/internal/pyneats_units_audit.md.