reach.length is 100x slower on EPANET res1d files - likely a mikeio1d exception path

Follow-up from #679.

`Res1DReach.__init__` reads `reach.length`, which goes to mikeio1d's
`result_reach.py:_get_total_length`. A test user profiled it at ~2.4 s over 8,377 calls on an
EPANET-backed file — ~286 µs per call. On `tests/testdata/network.res1d`, a MIKE urban model, the
same property costs 2.6 µs. A 100x gap.

The mikeio1d source is:

```python
def _get_total_length(self) -> float:
total_length = 0
try:
for reach in self.reaches:
total_length += reach.Length
return total_length
except Exception as _:
return total_length
```

The docstring says it returns zero when no length is available. So the likely explanation is that
`Length` is unavailable on EPANET links, every call raises, and the cost is exception handling
across the pythonnet boundary — which would also mean every reach gets length 0 and the graph edge
lengths are meaningless on those files.

Next step: confirm what `reach.length` returns on an EPANET link. If it is 0, this belongs upstream
in mikeio1d rather than worked around in modelskill. Deriving length from
`end_chainage - start_chainage` is tempting but untested and may hit the same path.

Topology reuse (#682) would avoid the cost in the calibration-loop case regardless, since the reach
loop never runs.

Start at Res1DReach.__init__ and inspect mikeio1d result_reach.py:_get_total_length; profile or read reach.length for an EPANET link and compare it with tests/testdata/network.res1d. Check whether the value is zero and whether the exception path crosses pythonnet, then determine whether the fix belongs upstream in mikeio1d or can be safely handled here; topology reuse in #682 is a relevant comparison.

Written by the indexing model from the issue text.