This corrects for compilers where int is too small to allow constants to be shifted as much as they are in the SECP256K1_FE_CONST_INNER macro.
On my x86_64 system this change results in a bit-for-bit identical test binary, with both 32- and 64-bit field and scalars.