Optimized fixed ints based on Cython

[psrok1]

Current design is based on metaclasses, int-based classes etc. All of these things slow down object construction which is crucial for performance (see also #95)

• Redesigned IntType to simplify objects (FixedInt). Changes are breaking, but we need to test it using existing modules to check if we don't break it too much.
• Added Cython-based implementation although it introduces limitations. For example:

      def __or__(self, FixedInt other):
          return FixedInt(self.longval | other.longval, self.bits, self.sign)
  
      def __or__(self, unsigned long long other):
          return FixedInt(self.longval | other, self.bits, self.sign)
  

  accepts only int that are within unsigned 64-bit range (no negative numbers) and arbitrary FixedInts
  On the other hand long long accepts only signed 64-bit range.
• Added Python-based implementation for non-2's complement architectures that is opt-in via MALDUCK_PYTHON_INTS environment variable or when import of native extension fails.

TODO:

• Consider renaming malduck to src. During local development, Python imports package from local directory and fallbacks to Python implementation when .so is not in malduck/ints directory.
• Modify CI/CD to prepare wheels for x86 glibc/musl (also check other architectures in qemu)
• Make better tests because some masking/sign extension operations are implemented by ourselves.