feat: add MPS-aware device pipeline and improve ASR timestamp handling

fix(docker): add setuptools to CPU requirements to resolve pkg_resources import error

refactor: simplify speaker diarizer pipeline setup

fix long audio failure test

fix(docker): add nvcc for FlashInfer JIT compilation on older GPUs

refactor(deploy): use CUDA_VISIBLE_DEVICES for multi-GPU instance control

feat(deploy): unify internal nginx startup for cpu/gpu and auto multi-gpu

Revert "fix(asr): serialize full long-audio requests to prevent cross-request mixing"

fix(asr): serialize full long-audio requests to prevent cross-request mixing

fix(asr): serialize main inference to avoid cross-request mixing

refactor(auth): remove unused auth helper

refactor(auth): centralize token extraction and validation

fix(auth): treat empty API_KEY as optional and support ws query token

refactor(audio-splitter): merge only short segments after VAD

tune(audio-splitter): reduce max silence merge threshold to 2s