[Codegen][CUDA] Fix make_int4x cuda codegen vectorize

[wyc-ruiker]

Added support for int4x32 int4x16 int4x4 in BroadcastNode.

In the int4x4 testcase, the IR is:

primfn(compute_1: handle) -> ()
  attr = {"global_symbol": "main", "tir.noalias": True}
  buffers = {compute: Buffer(compute_2: Pointer(int4), int4, [64, 4], [])}
  buffer_map = {compute_1: compute} {
  attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 64;
  compute_2[ramp((blockIdx.x*4), 1, 4)] = broadcast(1i4, 4)
}

Before the fix in codegen_c.cc, the codegen cuda is:

extern "C" __global__ void make_int4x4_kernel0(int* __restrict__ compute) {
  ((int16_t*)(compute + ((((int)blockIdx.x) * 4)) / 8))[0] = (int16_t)4369;
}

For int16_t, this index (((int)blockIdx.x) * 4)) / 8 is a bug.
After the fix in codegen_c.cc, the codegen cuda is:

extern "C" __global__ void make_int4x4_kernel0(int* __restrict__ compute) {
  ((int16_t*)(compute) + ((((int)blockIdx.x) * 4)) / 4)[0] = (int16_t)4369;
}

Could you please help review this fix? @vinx13 @Hzfengsy

[tqchen]

@vinx13 please help to manage this PR