[Perf] Avoid unnecessary cuMemcpy in to_numpy()

[xumingkuan]

Concisely describe the proposed feature
to_numpy() is sometimes too slow now on GPUs. This is because to_numpy() calls tensor_to_ext_arr(), which doesn't have the information that we don't need the original value of the newly created numpy array. So we have an unnecessary copy from host to device here:

taichi/taichi/backends/cuda/codegen_cuda.cpp

Lines 71 to 72 in c824a8b

	CUDADriver::get_instance().memcpy_host_to_device(
	(void *)device_buffers[i], host_buffers[i], args[i].size);

We can remove this CUDA memcpy to accelerate it.

Describe the solution you'd like (if any)
Add a compiler hint to tensor_to_ext_arr() when called by to_numpy() (or to_torch()), that we will fully write the values in the external array. Then we can remove the unnecessary copy in codegen_cuda.cpp.

Additional comments
See also ti.loop_unique(covers=...), but that one only supports SNodes now.

Shall we add the new hint to the CHI IR or to the kernel?

[k-ye]

Sounds great. Note that GLSL has type qualifiers like in, out or inout, which is a more generic approach than hinting tensor_to_ext_arr(). We can also analyze each kernel itself to see if the kernel only reads the external array, or does both read and write to it.

[bobcao3]

Merging with #4048