Avoid signed overflow in RGBLINK's +, -, and * - #2060

RGBLINK's RPN evaluator computed RPN_ADD, RPN_SUB, and RPN_MUL directly on int32_t, at src/link/patch.cpp lines 117, 121, and 124. Signed overflow is UB in C++, so any link-time expression whose result leaves the 32-bit signed range was undefined. RGBASM already handles this: Expression::makeBinaryOp in src/asm/rpn.cpp cast both operands to uint32_t for exactly those three operators and cast the result back. RGBLINK never got the same treatment.

The develop preset turns on UBSan, so it shows up right away:

$ cat over.asm
def fzero equs "startof(\"test\")"
section "test", rom0
dl $7fff_ffff + ({fzero} + 1)
$ ./rgbasm -o over.o over.asm && ./rgblink -o over.gb over.o
src/link/patch.cpp:117:26: runtime error: signed integer overflow: 1 + 2147483647 cannot be represented in type 'int32_t' (aka 'int')

I moved the three operators into opmath.cpp as op_add, op_sub, and op_mul, next to op_neg and the others that are already there to avoid C++'s behavior, and called them from both RGBASM and RGBLINK, so the two evaluators cannot drift apart again. RPN_ADD and RPN_MUL still pop both operands in one expression, which stays safe because both operations are commutative.

Values are unchanged on every platform RGBDS supports, so no existing expected output moved. The new test/link/patch-wraparound test covers all three operators: its .out.bin pins the wrapped results $80000000, $7FFFFFFF, and $7FFFFFFD, and its empty .out fails without the fix, because the three UBSan reports land on stderr.

Built with cmake -B build --preset develop and cmake --build build -j 4 on macOS 26 with Apple clang 17, then ran test/run-tests.sh --only-internal: 864 RGBASM, 115 RGBLINK, 441 RGBFIX, and 245 RGBGFX tests pass. I did not run the external-project tests.