Mesaure memory usage of decimal64_t operations

[mborland]

The default stack size on CUDA is 1024, which is blown by simple operations like addition. Maybe we have some space for optimization, or maybe we simply need to be more aggressive with forced inline on CUDA.

[mborland]

Decimal64_t basic ops need a stack of 4k, and decimal128_t basic ops need a stack of 8k