plot: fix limits handling, unknown kwargs and a old docstring reference

[janfb]

Follow-up to #1425, which PR #1631 largely delivered. These are the concrete items that survive, each small and independent:

1. prepare_for_plot compares limits == [], which is wrong for a np.ndarray. Use a length check, and align the pairplot / marginal_plot type hints. Needs a regression test.
2. Warn or raise on unrecognised keys in diag_kwargs, upper_kwargs and lower_kwargs. A typo is currently silent.
3. sbi/analysis/plot.py:1336 points readers at _get_default_fig_kwargs(), and the deprecation message at plot.py:2862 tells them to use it. That function does not exist. Only _get_default_opts is defined, at plot.py:2860.

[BHARATH0153]

all three items are real and reproducible on main. I worked through each one and have fixes ready (with a regression test for each)

1. prepare_for_plot compares limits == []

plot.py:884 currently does if limits is None or limits == []:. For a np.ndarray this is elementwise comparison, so passing a valid numpy array raises a cryptic error before any plotting happens:

ValueError: operands could not be broadcast together with shapes (2,2) (0,)
(a torch.Tensor is broken too — tensor == [] returns a tensor and the truth test fails). Fix is a length check, which works for list, np.ndarray and torch.Tensor alike:

if limits is None or len(limits) == 0:
limits = infer_limits(samples, dim, points)
I also aligned the public type hints so this is actually advertised: pairplot and marginal_plot now declare limits: Optional[Union[List, torch.Tensor, np.ndarray]] (matching what prepare_for_plot already accepts). The existing [] → infer behavior is preserved.

2. Silent unknown keys in diag_kwargs / upper_kwargs / lower_kwargs

A typo such as diag_kwargs={"bw_metod": "scott"} was silently absorbed by the deep-merge update() in _prepare_kwargs and then ignored. I added _validate_plot_kwargs, called from _prepare_kwargs, which raises a ValueError naming the offending keys and listing the valid ones, e.g.:

Unrecognised keyword(s) ['bw_metod'] in diag_kwargs for plot type 'hist'.
Valid options are ['bin_heuristic', 'mpl_kwargs'].
The valid set is derived from the *Options dataclasses (dataclasses.fields), i.e. it is the union of documented keys per plot family rather than the current plot type. That deliberately keeps cases like lower_kwargs=HistOffDiagOptions() with lower=None working (which the existing test_plotting_kwargs_validation treats as valid), while still catching typos — which are the actual silent-failure case in the issue.

3. Stale _get_default_fig_kwargs references

Both places that point readers at the non-existent _get_default_fig_kwargs() now reference FigOptions, which is the real source of truth for figure defaults (and what fig_kwargs accepts via _prepare_fig_kwargs):

the _arrange_grid docstring for fig_kwargs (plot.py:1336),
the deprecation message in _get_default_opts (plot.py:2862).

New regression tests in tests/plot_test.py: prepare_for_plot with list / np.ndarray / torch.Tensor limits; empty-list inference; pairplot + marginal_plot with numpy limits (asserting the xlim is applied); and ValueError on unknown keys in all three kwargs arguments.

[janfb]

item 1 is now fixed via #2000. thanks @BHARATH0153 for proposing it here and @dhruvb2028 implementing.