Documentation: reproducibility and numeric precision

[sjdv1982]

For reproducibility, it is essential:

• To perform Numpy calculations in float64. Matrix multiplication gives significant irreproducible errors otherwise.
• To return results in float32. The least significant bits contain random rounding noise that will change checksums. Casting to float32 removes this.