fix: ensure raveled array uses float64 dtype in DictToArrayBijection (closes #8337)

[botbikamordehai2-sketch]

What

When a PyMC model uses a length-one dimensioned vector RV as a non-sequence in pytensor.scan, the raveled NUTS setup fails because DictToArrayBijection.map does not enforce a consistent float64 dtype. The original code used np.concatenate on raveled arrays that may have incompatible dtypes (e.g., object from a 1-element vector), causing downstream errors.

Fix

Forced the raveled arrays to np.float64 before concatenation, and initialized the empty array with dtype=np.float64. This ensures the bijection always works with the expected numeric type.

Closes #8337

[read-the-docs-community]

Documentation build overview

📚 pymc | 🛠️ Build #33251418 | 📁 Comparing a26c6ba against latest (94787c0)

  🔍 Preview build  

2 files changed

± glossary.html
± _modules/pymc/blocking.html

[ricardoV94]

Users nay be trying to sample in float32 or all variables could be discrete. upcast dtype should be inferred from input types not forced

[ricardoV94]

Also the issue explanation seems doubtful. Have you shown it?