Fix dens(): honour p.rope and fail loudly on degenerate posteriors

[mjuez]

Summary

Two bugs found while helping a user diagnose a confusing dens() plot with LOOCV:

1. dens() ignored p.rope for the orange ROPE axvlines. They were hardcoded at ±0.01. If the user passed rope=0.05 or rope=0.001 to CorrelatedTTest, the plot showed lines at ±0.01 while posterior.probs() was correctly computed against the real rope — silently misleading.
2. dens() crashed with ValueError: Axis limits cannot be NaN or Inf from inside matplotlib when the posterior was degenerate (var=0, typical when the two score arrays are identical). The user got a numpy/matplotlib stack trace with no hint of the cause.

Fixes

axvline honours p.rope

```diff

• ax.axvline(0.01, c="darkorange", linewidth=2, zorder=101)
• ax.axvline(-0.01, c="darkorange", linewidth=2, zorder=101)

• ax.axvline(p.rope, c="darkorange", linewidth=2, zorder=101)
• ax.axvline(-p.rope, c="darkorange", linewidth=2, zorder=101)
  ```

Explicit error on degenerate posterior

```python
if not (p.var > 0):
raise ValueError(
f"Cannot draw density: posterior variance is {p.var!r}. "
"The two classifiers likely produced identical scores; "
"the posterior is degenerate and has no density to plot."
)
```

The check covers var=0, var<0 (shouldn't happen but defensive), and var=NaN. Users get an actionable message instead of a leaked traceback from matplotlib.

Tests

• TestRopeAxvline — parametrised over rope ∈ {0.001, 0.01, 0.05}, asserts the two ROPE axvlines sit at exactly ±rope.
• TestDegeneratePosterior — asserts ValueError (with \"degenerate\" in the message) when the posterior has var=0.

Compatibility

• No public API changes. dens() still accepts the same arguments and returns the same figure object.
• Behaviour change is strictly an improvement: same input → correct visual output (case 1), degenerate input → clear error instead of confusing traceback (case 2).
• Existing pytest-mpl baselines reproduce identically because the original tests use rope=0.01, which is what was hardcoded — the fix is a no-op for that case.

Bumped to 1.2.1.

Test plan

• pytest --mpl — 23/23 passing locally.
• python -m build — builds wheel + sdist.
• twine check dist/* — PASSED.
• CI matrix (will run on this PR).
• After merge: tag v1.2.1 → publish to PyPI via the existing release workflow.

🤖 Generated with Claude Code