prod_env_mat: allocate GPU memory out of frame loop (#2832)

Allocating GPU memory is not a cheap operator. This PR allocates memory
for `int_temp`, `uint64_temp`, and `tensor_list[0, 1, 3, 4, 5, 6]` out
of the frame loop, so they can be reused in each loop without allocating
many times.
In the original code, `tensor_list[3]`, `tensor_list[4]`, and
`tensor_list[6]` may need to reallocate if the memory is not enough.
This behavior still exists.
The shape of `tensor_list[2]` is dynamic, so it is not refactored in
this PR.
With CUDA enabled, unit tests for C++ and Python can pass. The examples
can be performed.
The speedup can be observed when the number of frames (samples) in a
batch is not small.

---------

Signed-off-by: Jinzhe Zeng <[email protected]>