Preserve Gauss-von-Mises single-point autograd

[FlorianPfaff]

Bug

GaussVonMisesDistribution.pdf(...) kept backend-native values for batched evaluation, but the single-point path returned float(p[0]).

On the PyTorch backend, a query tensor with requires_grad=True therefore produced a Python float. The conversion silently detached the density from the computation graph, so callers could not differentiate a single-point likelihood even though the equivalent batched path remained differentiable.

Fix

Return the backend scalar p[0] directly for single-point evaluation instead of converting it to Python float.

This preserves the existing numerical value while keeping the active backend's dtype, device, and autograd history.

Regression coverage

Add a PyTorch-backend regression that evaluates a single Gauss-von-Mises point with requires_grad=True and verifies:

• the result is a scalar PyTorch tensor;
• dtype and device match the query;
• the density is finite and positive;
• backpropagation reaches both query coordinates with finite, nonzero gradients.

Validation

• independently confirmed that converting a differentiable PyTorch scalar with float(...) returns a plain Python float and loses autograd;
• final production diff is one line;
• branch is based directly on current main (f196ec41d1d8681e279936db9f1a0a3abfa276e7) and is 2 commits ahead / 0 behind;
• GitHub Actions is authoritative for the full backend, lint, packaging, documentation, and integration matrix.