fix: add native RTX 50xx (Blackwell / sm_120) GPU support

[NotYuSheng]

Closes #194
Relates to #179

Summary

• Switch backend base image from pytorch:2.12.0-cuda12.6 to pytorch:2.8.0-cuda12.8-cudnn9-runtime — first PyTorch release with native sm_120 (Blackwell) compiled kernels; the previous image only covered up to sm_90
• Pre-load audio via torchaudio.load and pass {"waveform": ..., "sample_rate": ...} to the pyannote pipeline instead of a file path, bypassing torchcodec which fails under CUDA 12.8
• CUDA 12.x preserves libcublas.so.12, so ctranslate2/faster-whisper is unaffected

Test plan

• Build the Docker image and confirm no dependency errors
• Run diarization on a machine with an RTX 50xx GPU (sm_120) to confirm the CUDA kernel error is resolved

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the backend Dockerfile to use PyTorch 2.8.0 with CUDA 12.8 and modifies the diarization service to pre-load audio via torchaudio to bypass torchcodec compatibility issues. However, the synchronous torchaudio.load call is executed on the main event loop, which will block concurrent requests. It is recommended to offload both the audio loading and pipeline execution to the executor.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.