Fix CUDA detection and modern FFmpeg compatibility

[DrStone1971]

This PR updates Decord to build on newer Linux/CUDA/FFmpeg environments.

Changes:

• Replace legacy CUDA detection with CMake CUDAToolkit
• Fix CUDA detection with CUDA 12/13
• Add modern FFmpeg bitstream filter header usage
• Fix FFmpeg const-correctness issues with AVCodec and AVInputFormat
• Replace deprecated FFmpeg rotation side-data access for newer libavformat
• Fix C++ warnings such as catching dmlc::Error by value
• Add bounds check for CUDA codec string lookup

Tested on:

• Ubuntu Linux
• GCC 13
• CUDA 13.2
• CMake + Ninja
• FFmpeg / LibAV: libavcodec 60, libavformat 60, libavutil 58
• Python 3.12
• USE_CUDA=ON
• CMAKE_CUDA_ARCHITECTURES=120

Also verified local wheel build and import successfully.

[DrStone1971]

Tested on modern environment:

• Ubuntu 24.04
• GCC 13
• CUDA 13.2 (Blackwell sm_120)
• CMake + Ninja
• FFmpeg (libavcodec 60 / libavformat 60)
• Python 3.12

Build completed successfully with USE_CUDA=ON.
Wheel package builds and imports correctly.

This removes the need for -fpermissive and fixes compatibility with newer FFmpeg APIs.