Fix compute_marginals and sample_posterior for plates of size 1

[MohammadHijjawi97]

Fixes #3102

Proposed changes

TraceEnum_ELBO.compute_marginals() and TraceEnum_ELBO.sample_posterior() raise KeyError when an enumerated site is inside a plate of size 1. Packed tensors drop size-1 dims, so the plate's dim never appears in any factor; contract_to_tensor() was called with only a cache, so its default LogRing had an empty dim_to_size and Ring.broadcast() could not look up the plate size when broadcasting the result back to the site's ordinal.

This passes a LogRing whose dim_to_size is filled from the site's plate frames, in both _compute_marginals and BackwardSampleMessenger.

Tests

• New test_marginals_plate_3102[1,2,3] in tests/infer/test_enum.py checks the marginals against the analytic posterior and runs sample_posterior; size=1 fails with KeyError: 'a' before this change. The script from the issue now prints tensor([0.3000, 0.7000]).
• pytest tests/infer/test_enum.py (678 passed) and pytest tests/infer/test_discrete.py.
• ruff check / ruff format --check / mypy on the changed files.