C codegen with Target::CPlusPlusMangling generates incorrect type cast

[knzivid]

Steps

1. Build release/13.x halide python bindings
2. Run the python snippet at the end
3. Compile the output C file with g++ into an object

Expected: Since TargetFeature.CPlusPlusMangling is enabled, I can compile the generated C source with g++.

Actual: Compile error. I extracted the offending code to https://www.godbolt.org/z/bTbT7PKaT

reprod.c: In function ‘int consumer(halide_buffer_t*)’:
reprod.c:2657:47: error: invalid conversion from ‘const void*’ to ‘void*’ [-fpermissive]
 2657 |  void *const _60 = ((const void * *)_opencl)[0];
      |                    ~~~~~~~~~~~~~~~~~~~~~~~~~~~^
      |                                               |
      |                                               const void*
reprod.c:2897:49: error: invalid conversion from ‘const void*’ to ‘void*’ [-fpermissive]
 2897 |   void *const _202 = ((const void * *)_opencl)[0];
      |                      ~~~~~~~~~~~~~~~~~~~~~~~~~~~^
      |                                                 |
      |                                                 const void*

Here are the two offending snippets

 void *const _56 = _halide_buffer_get_host(_opencl_buf_buffer);
 void * _opencl = _56;
 void *const _57 = _halide_buffer_get_host(_opencl_gpu_source_kernels_buffer);
 int32_t const _58 = halide_opencl_initialize_kernels(_ucon, _opencl, _57, 4166);
 bool const _59 = _58 == 0;
 if (!_59)
 {
  return _58;
 }
 void *const _60 = ((const void * *)_opencl)[0]; // error 1

  void *const _202 = ((const void * *)_opencl)[0]; // error 2

---

def find_gpu_target():
    host_target = get_host_target()
    return Target(host_target.os, host_target.arch, host_target.bits,
                  [TargetFeature.OpenCL, TargetFeature.CLDoubles, TargetFeature.CPlusPlusMangling])

x, y = Var(), Var()
consumer = Func("consumer")
lut_const = np.arange(128)
lut_buffer = Buffer(lut_const)

at_lut = Func("at_lut")
i = Var()
at_lut[i] = lut_buffer[i] % 16
consumer[x, y] = at_lut[(x + y) % lut_const.shape[0]] * (x * x + y)

xo, yo, xi, yi = Var(), Var(), Var(), Var()
consumer.gpu_tile(x, y, xo, yo, xi, yi, 8, 8)

consumer.compile_to(outputs={
    Output.c_source: str(Path(__file__).parent / "reprod.c"),
}, arguments=[], fn_name="consumer", target=find_gpu_target())

[steven-johnson]

fixed by #6638