Improve conda + pip build instructions in README.md and Sphinx documentation.

[ma595]

Deploying detectron2 easily across all systems is still not well documented. I propose a combined conda + pip approach as below:

conda install -c conda-forge gdal
conda install pytorch torchvision torchaudio pytorch-cuda=<cuda-version> -c pytorch -c nvidia
pip install .[ci] 

The first line ensures that the project can be installed on systems lacking the GDAL headers. The second conda line does the same, but for CUDA. pytorch binaries shipped with pip also come with their own CUDA runtime (pip install torch defaults to 12.1 as of 02/12/23). The only requirement for the latter to work is that global CUDA is >= 9.0.

Investigate whether including pytorch dependencies in setup.py or pyproject.toml is the right approach, or whether it is best to defer to the user. Either way, this needs to be commented, as we are relying on the outdated detectron2 instructions