Fix overflow checking for operations with mixed sign

[waj]

This is a rebuilt of arithmetic operations with overflow checking (+, -, *) to fix issues on cases with arguments of different sign.

I implemented the same approach used by clang (https://github.com/llvm/llvm-project/blob/796898172c48a475f27f038e587c35dbba9ab7a6/clang/lib/CodeGen/CGBuiltin.cpp#L3378-L3465) for the builtin operations (__builtin_add_overflow, etc.) and using the same workaround for multiplication of integers mixed sign. This workaround is needed to avoid internal i128 multiplication for i64 values and to avoid sigfaults for i128 multiplications.

This also adds specs to test operations between all integer types, using some extreme values from each type and contrasting results using BigInt.

Before implementing this approach I've been testing a port of SafeInt. It's a more exhaustive case by case approach that produces shorter binary code, but it requires more work and detailed testing. I assume shorter also means faster although I didn't perform any benchmark. I decided to go with this easier but correct code first.

Fixes #9277

[RX14]

Is it correct that this uses non-power-of-two LLVM types like i33?

[bcardiff]

Yes, since LLVM 2.5. Spooky 👻

The LLVM native code generator now supports arbitrary precision integers. Types like i33 have long been valid in the LLVM IR, but were previously only supported by the interpreter. Note that the C backend still does not support these.
https://releases.llvm.org/2.5/docs/ReleaseNotes.html

[asterite]

Hmm... yeah, like i128 is "supported". But I guess if CI passed then everything's fine :-)

[waj]

i128 is supported but it requires compiler-rt. In macOS normaly clang is used during linking and it makes it automatically available. That's why I enabled Int128 and UInt128 tests only in darwin. We need to fix it but that's a separate issue.