Fix dtype casting error in scaled_centered_kpi method

[ishanrajsingh]

Fix dtype casting error in AKS algorithm when enable_aks=True

• Convert kpi and population arrays to float64 in scaled_centered_kpi method
• Prevents UFuncTypeError: Cannot cast ufunc 'divide' output from
  dtype('float64') to dtype('int64') with casting rule 'same_kind'
• Use explicit dtype=np.float64 for output arrays in np.divide operations
• Maintains numerical accuracy while handling integer input data
• Enables successful prior and posterior sampling with enable_aks=True
• Also documents TensorFlow dtype compatibility fix in comments

Fixes #1042

Tested with:

• Integer KPI data (int32) that previously caused casting errors
• Both national and geo-level models
• Prior sampling with enable_aks=True
• Posterior sampling with enable_aks=True
• Maintains backward compatibility when enable_aks=False