fix: use float64 to avoid truncation (#100)

* fix: use float64 to avoid truncation

* explanation

src/ess/nmx/scaling.py

@@ -359,6 +359,10 @@ def polyval_wavelength(
 
 
     """
+    # We need to use float64 precision because
+    # the curve fit routine depends on finite difference
+    # estimates of the derivative.
+    wavelength = wavelength.to(dtype='float64')
     out = sc.zeros_like(wavelength)
     out.unit = out_unit
     xk = sc.ones_like(wavelength)