Fix signed overflow in modem training phase arithmetic

[coral]

Ran into an interesting issue where compiling with GCC at -O2 and -O3 made modem training unstable but compiling the receiver at -O0, or with -fwrapv, made it work. Eventually tracked it down to this.

The phase angles represent a full turn in 32 bits, so the differences need to wrap around. For example, going from 0x7ffffff0 to 0x80000010 should give a difference of 32, but subtracting those as signed integers overflows. Casting the result to unsigned afterward is too late.

This moves the casts before the arithmetic so the wraparound is defined, then casts back to signed where the training logic expects it.