add Tracker protocol and training tracking guide.

[janfb]

Context

We want a simple, extensible way to log training metrics from sbi that works with
TensorBoard by default while letting users plug in their own experiment trackers.
The previous summary_writer argument was TensorBoard-specific and hard to extend.

• Offer a minimal protocol for logging metrics that supports external tools.
• Keep default behavior unchanged (TensorBoard logs with a standard log directory).
• Make it easy to add third-party trackers without changing trainer internals, it all happens on the user side.

What changed

• Introduced a Tracker protocol (sbi/sbi_types.py) with metric/figure/param logging and a log_dir property.
• Added TensorBoardTracker in sbi/utils/tracking.py as the default implementation.
• Trainers now accept tracker= and log through the protocol; summary_writer= is a deprecated alias that wraps TensorBoard and warns if used.
• plot_summary now reads log_dir from the tracker and errors with a clear message if a non-TensorBoard tracker is used.
• add experiment tracking how-to-guide; the tracking guide includes a runnable TensorBoard example, adapter patterns for W&B/MLflow/Trackio.

Naming choices

• Tracker: short and tool-agnostic; aligns better with MLflow/W&B conventions
  than “writer”.
• add_figure: mirrors TensorBoard’s SummaryWriter.add_figure, keeping the
  default implementation a thin adapter.

Notes

• External trackers are intentionally not bundled; users can implement small adapters
  that satisfy the Tracker protocol.
• summary_writer remains for backward compatibility but is deprecated.

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
5816	1	5815	146

View the full list of 1 ❄️ flaky test(s)

tests/torchutils_test.py::TorchUtilsTest::test_searchsorted

Flake rate in main: 42.53% (Passed 50 times, Failed 37 times)

Stack Traces | 0.005s run time

.venv/lib/python3.10....../site-packages/xdist/remote.py:289: in pytest_runtest_logreport
    self.sendevent("testreport", data=data)
.venv/lib/python3.10....../site-packages/xdist/remote.py:126: in sendevent
    self.channel.send((name, kwargs))
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:912: in send
    self.gateway._send(Message.CHANNEL_DATA, self.id, dumps_internal(item))
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1629: in dumps_internal
    return _Serializer().save(obj)  # type: ignore[return-value]
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1647: in save
    self._save(obj)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1667: in _save
    dispatch(self, obj)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1744: in save_tuple
    self._save(item)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1667: in _save
    dispatch(self, obj)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1740: in save_dict
    self._write_setitem(key, value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1734: in _write_setitem
    self._save(value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1667: in _save
    dispatch(self, obj)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1740: in save_dict
    self._write_setitem(key, value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1734: in _write_setitem
    self._save(value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1667: in _save
    dispatch(self, obj)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1740: in save_dict
    self._write_setitem(key, value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1734: in _write_setitem
    self._save(value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1667: in _save
    dispatch(self, obj)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1740: in save_dict
    self._write_setitem(key, value)
.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1734: in _write_setitem
    self._save(value)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <execnet.gateway_base._Serializer object at 0x7ff4b46cd420>
obj = tensor([0.0000, 0.1111, 0.2222, 0.3333, 0.4444, 0.5556, 0.6667, 0.7778, 0.8889])

    def _save(self, obj: object) -> None:
        tp = type(obj)
        try:
            dispatch = self._dispatch[tp]
        except KeyError:
            methodname = "save_" + tp.__name__
            meth: Callable[[_Serializer, object], None] | None = getattr(
                self.__class__, methodname, None
            )
            if meth is None:
>               raise DumpError(f"can't serialize {tp}") from None
E               execnet.gateway_base.DumpError: can't serialize <class 'torch.Tensor'>

.venv/lib/python3.10....................................................../site-packages/execnet/gateway_base.py:1665: DumpError

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}

[manuelgloeckler]

Thats looks great!

Also nice that one can now e.g. easily switch to e.g. wandb.