fix: import numpy for eval() of xlsx columns under numpy >=2.0

[wangliangliang2008]

Problem

On numpy 2.0+, pandas writes numeric values to xlsx using the new repr format
(np.float64(0.355) instead of 0.355). Strings in the lines and
new_sub_times columns of output/audio/tts_tasks.xlsx therefore look like:

'[[np.float64(5.205), np.float64(7.965)]]'

These columns are deserialized via eval() in:

• core/_10_gen_audio.py (lines 68, 165, 187)
• core/_11_merge_audio.py (lines 19, 22, 32)

Neither file imports numpy, so eval() raises:

NameError: name 'np' is not defined

In the Streamlit UI this surfaces as ❌ 任务出错: name 'np' is not defined
during the dubbing step, blocking the whole pipeline at the audio-merge stage.

requirements.txt already pins numpy>=2.0.2, so any fresh install will hit
this consistently.

Reproduction

1. Fresh install on macOS / Linux (or anywhere numpy>=2.0 is installed)
2. Run any video through transcription → translation → TTS → dubbing
3. At the "Audio merging" step in tab c. 配音, the task aborts with the
   NameError above

Fix

Add import numpy as np to both files so eval() finds it in scope. Two
files, two lines + an explanatory comment.

Test

Tested on macOS 26.5 + Python 3.10.20 + numpy 2.2.6, with a 4-minute YouTube
clip translated and dubbed end-to-end. Pipeline now completes through:

• output/dub.mp3
• output/normalized_dub.wav
• output/dub.srt
• final merged video

Related

When testing this I also noticed install.py doesn't detect whether the user's
ffmpeg was built with libass. The macOS Homebrew lite ffmpeg formula
ships without libass, which means the subtitles= filter is missing entirely
and the final video merge fails with cryptic "No option name near 'output/...'"
errors. Worth a separate issue / PR to make install.py warn — happy to file
one if interested.