BETA CUDA interface: Fix CUDA context initialization

[NicolasHug]

On main, this fails:

from torchcodec.decoders import VideoDecoder
from joblib import Parallel, delayed

video_path = "/home/nicolashug/videos_h264/vid.mp4"

import torch

def decode_one_video():
    decoder = VideoDecoder(video_path, device="cuda:0:beta", seek_mode="approximate")
    decoder.get_frame_at(-1)

Parallel(n_jobs=8, prefer="threads")(delayed(decode_one_video)() for _ in range(100))

with

RuntimeError: Failed to get decoder caps: 201

That is, spawning one VideoDecoder per thread fails with CUDA error 201 which is CUDA_INVALID_CONTEXT.

Uh?

I'm confused as well. This seems to indicate that our CUDA context initialization hack, where we create a dummy tensor to force context creation, doesn't work as expected:

torchcodec/src/torchcodec/_core/BetaCudaDeviceInterface.cpp

Lines 206 to 208 in 9c5da20

	// Initialize CUDA context with a dummy tensor
	torch::Tensor dummyTensorForCudaInitialization = torch::empty(
	{1}, torch::TensorOptions().dtype(torch::kUInt8).device(device_));

After a lot of trial and error, it seems that using torch::zeros instead of torch::empty resolves the problem. Why? I have no idea. Maybe torch::empty was optimized out? Maybe, but that doesn't explain why the default CUDA interface works fine with teh snippet above... Anyway, now they both use torch::zeros, and they both work when running my multithreaded benchmarks.