Fix cross-platform negative value conversion for unsigned integer types

[kojix2]

Fixed test failures on macOS.

Problem: casting negative float to unsigned int works differently on macOS.

• Linux: result is a large number (two’s complement).
• macOS: result is 0 (unexpected).

Solution: updated macros to handle negative values the same on all systems:

#define m_from_double(x) ((x) < 0 ? (dtype)((long long)(x)) : (dtype)(x))

[kou]

We need #225 before this.