Fix a rare CUDA compile error.

Despite having compiled this CUDA code on many different machines, only
ozstar threw a compile error surrounding taking the absolute value of
an int8_t m_accum value and adding/subtracting it with a float. This is
fixed by pre-computing the absolute values of these m_accums and passing
them in. There appears to be no impact on performance.

While we're here, I made more pointers const. This unfortunately also
did not affect performance; maybe the compiler was already good about
doing associated optimisations.

Also remove a directive that ignored null pointer warnings; the
newest version of bindgen has fixed things.

Also tidy the body of fee_kernel; I like not having to indent code if I
don't have to.

CHANGELOG.md

@@ -8,6 +8,10 @@ Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 ### Fixed
+- CUDA compilation on ozstar failed because of an arithmetic operation between
+  two different types. Compilation has succeeded elsewhere, such as on Ubuntu,
+  Arch, Pawsey's garrawarla and DUG. The code has changed to prevent the issue
+  in the first place and no compilation issues have been spotted.
 - CUDA function prototypes were being included in the C header, even if no CUDA
   feature was enabled.
 - The CUDA library libcudart was always statically linked by mistake. It is now