Boost.Multiprecision causes GPU (CUDA) build failures

[garth-wells]

Boost.Multiprecision has some bugs that causes build failures with nvcc (CUDA compiler). This affects DOLFINx when building with nvcc, and when building code against DOLFINx using nvcc because the Boost.Multiprecision gets exposed in the public interface (it would be better if it didn't).

See boostorg/math#1285 and boostorg/multiprecision#695.

[jhale]

My understanding is that this is now fixed upstream.

Nonetheless, looking at https://github.com/FEniCS/dolfinx/blob/main/cpp/dolfinx/geometry/gjk.h it should be the consumer of the header that sets the (possibly extended precision) type U. Edit: This means we could remove boost-multiprecision as a dependency from C++, and add it only in Python.