Expand llvm.{u,s}mul.with.overflow before SPIR-V emission

[pvelesko]

Fixes #1391

Expands llvm.umul.with.overflow and llvm.smul.with.overflow into plain integer arithmetic in the post-link pass pipeline, so the construct never reaches SPIR-V.

Without this, clang emits llvm.umul.with.overflow.i64 for a device-side new T[n] or __builtin_mul_overflow, the SPIR-V producer turns it into a spirv.llvm_umul_with_overflow_i64 helper, and the consumer maps that back onto the intrinsic, renames the declaration to old_llvm.umul.with.overflow.i64 and leaves it undefined. Since chipStar puts all device code of a binary into one module, the whole program stops working.

Before and after on Aurora PVC (Intel Data Center GPU Max 1550), Level Zero, LLVM 22.1.3 with the in-tree SPIR-V backend, running the new regression test:

### BEFORE
  case 0: got lo=0 ov=0, want lo=42 ov=0
  case 1: got lo=0 ov=0, want lo=18446744073709551614 ov=1
  case 3: got lo=0 ov=0, want lo=0 ov=1
FAIL

CHIP error : hipErrorInvalidImage (Level Zero module build failed:
error: undefined reference to `old_llvm.umul.with.overflow.i64'
in function: 'old_llvm.umul.with.overflow.i64' called by kernel: 'testUmulOverflow(unsigned long*, unsigned long*)'
error: backend compiler failed build.

### AFTER
PASS

Counting umul_with_overflow in the extracted SPIR-V of the same source goes from 7 to 0.

The overflow predicate uses a division rather than a 128-bit multiply, since i128 is not portable across the SPIR-V consumers chipStar targets. These intrinsics come from allocation-size checks, so they are not hot. The signed case splits out a == -1 instead of folding it into the division, because sdiv INT_MIN, -1 is itself undefined.

The regression test uses __builtin_mul_overflow rather than new T[n], so that it exercises the intrinsic without also depending on the device heap, which has a separate problem.