Support SWMM by reading topology from the companion .inp file

SWMM results are refused today because the reach connectivity is missing from the `.out` file. It is not a mikeio1d gap — the data is not in the file, and it *is* in the companion `.inp` input file, which sits next to every SWMM result. Everything below was verified against the committed `tests/testdata/swmm.out` and the `swmm.inp` in [DHI/mikeio1d's testdata](https://github.com/DHI/mikeio1d/tree/main/tests/testdata).

## The topology is in the `.inp`, and it pairs exactly

`swmm.inp` is in the same upstream directory our four existing fixtures came from (MIT, commit `d937466`, recorded in `tests/testdata/README.md`), so it can be vendored the same way.

| | `swmm.out` | `swmm.inp` | match |
|---|---|---|---|
| reaches / links | 13 | 13 `[CONDUITS]` | exact |
| nodes | 14 | 13 `[JUNCTIONS]` + 1 `[OUTFALLS]` (`18`) | exact |

`[CONDUITS]` gives `Name From Node To Node Length`, so it supplies both connectivity *and* real reach lengths:

```
[CONDUITS]
;;Name From Node To Node Length Roughness ...
;;-------------- ---------------- ---------------- ---------- ----------
1 9 10 400 0.01
4 19 20 200 0.01
...
```

There is no fallback if the `.inp` is absent. In the `.out` alone, every reach reports `StartNodeIndex == -1`, node `xcoord`/`ycoord` are `nan`, `geometry` is `NodePoint(x=nan, y=nan)`, `chainages` is a single `-1e-30` sentinel, `length` is `0`, and `structures` is empty. So no geometric reconstruction is possible — the `.inp` is required, not optional.

Node and reach timeseries load fine, so only the topology is missing.

## Blocker: `_simplify_colnames` raises on every SWMM node

This has to be fixed before any of the above matters. mikeio1d reports each pollutant as `SWMM_NODE_QUAL` with no pollutant identity attached, so a node with two pollutants yields two identically named columns:

```python
>>> Res1D("swmm.out").nodes["9"].to_dataframe().columns
[..., 'SWMM_NODE_QUAL:9', 'SWMM_NODE_QUAL:9']
```

`_simplify_colnames` in `src/modelskill/model/adapters/_res1d.py` requires exactly one column per quantity, so it raises:

> ValueError: There must be exactly one column per quantity, found [('SWMM_NODE_QUAL:9', 'SWMM_NODE_QUAL'), ('SWMM_NODE_QUAL:9', 'SWMM_NODE_QUAL')]

The same happens for `SWMM_LINK_QUAL` on reaches.

The only source of the pollutant names is the `.inp` `[POLLUTANTS]` section, which `from_swmm` would already be reading. In the fixture it lists `TSS` then `Lead`, and `Lead` declares a co-pollutant fraction of `0.2` against TSS. The two columns hold `15.720075` and `3.144015` — exactly 0.2×, which confirms column order follows `[POLLUTANTS]` order *in this file*.

What is not established is whether that ordering is guaranteed across SWMM versions and models. That decides the naming:

- `SWMM_NODE_QUAL_TSS` / `SWMM_NODE_QUAL_Lead` — useful, but confidently wrong if the order ever differs
- `SWMM_NODE_QUAL_1` / `_2` — always correct, tells the user nothing

Worth checking the SWMM binary output format spec before choosing.

## Sketch

- `Network.from_swmm(out, inp, *, nodes=None, reaches=None)`, with `inp` required.
- Extend `src/modelskill/model/adapters/_inp.py` (added for EPANET in #687) with a `[CONDUITS]` reader. The section reader already handles `[SECTION]` headers, `;`-prefixed comments and whitespace-delimited rows.
- The reach adapter needs connectivity from the parsed rows rather than from `reach.start_node`. `Res1DReach` already accepts the start/end `Res1DNode` objects and only uses `reach.start_node` to validate them, and it gained a `length` override in #687 — so this is a narrower change than it looks.
- Other SWMM link types (`[PUMPS]`, `[ORIFICES]`, `[WEIRS]`, `[OUTLETS]`) all put `Name From-Node To-Node` in their first three columns and carry no length. The fixture has none of them, so supporting them would be unverified; decide whether to read them or refuse a model that uses them.
- `.out` moves out of `_UNSUPPORTED_EXTENSIONS` into its own extension set, and `_EXTENSION_CONSTRUCTORS` gains the `from_swmm` mapping. `test_every_mikeio1d_extension_is_accounted_for` keeps the tables honest.

## Related limits

- SWMM link quantities (`SWMM_LINK_FLOW`, `DEPTH`, `VELOCITY`, `CAPACITY`, …) sit on single-gridpoint reaches, which produce no breakpoints, so they are unreachable for the same reason EPANET's are — #680. Node quantities are unaffected, so a SWMM network would be usable via `NodeObservation` from day one.
- `[COORDINATES]` gives node x/y, but `NetworkNode` carries no coordinates today, so that is out of scope.
- MOUSE and Water Hammer remain refused for lack of any fixture — #686.

Supersedes #688, which claimed this was blocked on a fixture that turned out to exist.

Start with src/modelskill/model/adapters/_inp.py and _res1d.py, especially _simplify_colnames, then inspect the extension constructor tables and the committed tests/testdata/swmm.out and swmm.inp fixtures. Resolve how pollutant columns are named and which SWMM link sections are supported before implementing the required .inp topology path. Done means SWMM loading no longer raises for duplicate quality columns, topology and lengths come from the fixture, and the relevant extension-accounting tests pass.

Written by the indexing model from the issue text.