fix: 修改modal_infer.exe 为单独打包 + CI support

[Randomless]

背景

• 关联 运行错误，提示如下，什么问题 #11：反馈 v1.6 exe 仅输出 runtime hook 后退出，疑似打包/依赖冲突。
• 本 PR 需要把 project.spec改回v1.4/v1.5版本，保持 infer.exe 打包逻辑不变，先把 Modal CLI 打包流程独立为modal.spec并加 CI 校验，降低依赖冲突风险。
• modal.exe打包为onefile独立文件。

变更

• 新增 modal.spec：单文件打包 Modal 客户端，并打包 environment-cuda128.yml。
• 新增 environment-modal.yml：本地运行 Modal CLI 的轻量环境。
• build_windows.py：
  • infer 打包成功后自动构建 modal.spec进行打包；
  • 若当前环境缺少 modal/questionary，先自动安装；
  • modal_infer.exe 输出到 dist/faster_whisper_transwithai_chickenrice，并使用独立 build/modal 目录。
• modal_infer.py：
  • 强制 stdout/stderr 为 UTF‑8（修复 CI/Windows 下 argparse --help 的编码崩溃）；
  • 远程仓库版本固定为 v1.4（避免远端代码漂移）。
• CI：build-release-conda.yml 增加 modal_infer.exe --help 校验。

测试

• CI step（build-release-conda）需要检查两个exe的--help的内容是否正常输出，而不是空白：
  • infer.exe --help
  • modal_infer.exe --help

[Randomless]

还需要@grider-transwithai 改一下project.spec到老版本

错误的CI结果

Run cd dist\faster_whisper_transwithai_chickenrice
  cd dist\faster_whisper_transwithai_chickenrice
  infer.exe --help
  modal_infer.exe --help
  shell: C:\Windows\system32\cmd.EXE /C CALL {0}
  env:
    PYTHONUNBUFFERED: 1
    INPUT_RUN_POST: true
    CONDA: C:\Users\runneradmin\miniconda3
Runtime hook: Set KMP_DUPLICATE_LIB_OK=TRUE to resolve OpenMP conflicts
Runtime hook: Set TRANSFORMERS_NO_ADVISORY_WARNINGS=1 to suppress advisory warnings
Runtime hook: Configured ONNX Runtime to use 2 threads (half of 4 available CPUs)
usage: modal_infer.exe [-h] [--non-interactive]

feature-modal 运行脚本

options:
  -h, --help         show this help message and exit
  --non-interactive  跳过 questionary，全部通过参数指定（暂未实现）。

正确的CI结果

Run cd dist\faster_whisper_transwithai_chickenrice
Runtime hook: Set KMP_DUPLICATE_LIB_OK=TRUE to resolve OpenMP conflicts
Runtime hook: Set TRANSFORMERS_NO_ADVISORY_WARNINGS=1 to suppress advisory warnings
Runtime hook: Configured ONNX Runtime to use 2 threads (half of 4 available CPUs)
usage: infer.exe [-h] [--model_name_or_path MODEL_NAME_OR_PATH]
                 [--device DEVICE] [--compute_type COMPUTE_TYPE] [--overwrite]
                 [--audio_suffixes AUDIO_SUFFIXES] [--sub_formats SUB_FORMATS]
                 [--output_dir OUTPUT_DIR]
                 [--generation_config GENERATION_CONFIG]
                 [--log_level LOG_LEVEL] [--merge_segments]
                 [--no_merge_segments] [--merge_max_gap_ms MERGE_MAX_GAP_MS]
                 [--merge_max_duration_ms MERGE_MAX_DURATION_MS]
                 [--vad_threshold VAD_THRESHOLD]
                 [--vad_min_speech_duration_ms VAD_MIN_SPEECH_DURATION_MS]
                 [--vad_min_silence_duration_ms VAD_MIN_SILENCE_DURATION_MS]
                 [--vad_speech_pad_ms VAD_SPEECH_PAD_MS] [--console]
                 [--enable_batching] [--batch_size BATCH_SIZE]
                 [--max_batch_size MAX_BATCH_SIZE]
                 ...

Whisper transcription with custom VAD injection

positional arguments:
  base_dirs             Directories or files to process

options:
  -h, --help            show this help message and exit
  --model_name_or_path MODEL_NAME_OR_PATH
                        Path to the Whisper model
  --device DEVICE       Device to run the model on (cpu, cuda, auto)
  --compute_type COMPUTE_TYPE
                        Compute type for the model
  --overwrite           Overwrite existing subtitle files
  --audio_suffixes AUDIO_SUFFIXES
                        Comma-separated list of audio file extensions to
                        process
  --sub_formats SUB_FORMATS
                        Comma-separated list of subtitle formats (lrc, srt,
                        vtt, txt)
  --output_dir OUTPUT_DIR
                        Output directory for subtitle files
  --generation_config GENERATION_CONFIG
                        Path to generation config file
  --log_level LOG_LEVEL
                        Logging level
  --merge_segments      Enable segment merge post-processing (override config
                        file)
  --no_merge_segments   Disable segment merge post-processing (override config
                        file)
  --merge_max_gap_ms MERGE_MAX_GAP_MS
                        Max allowed gap (ms) between segments for merging
                        (override config file)
  --merge_max_duration_ms MERGE_MAX_DURATION_MS
                        Max duration (ms) of a merged segment (override config
                        file)
  --vad_threshold VAD_THRESHOLD
                        Override VAD threshold
  --vad_min_speech_duration_ms VAD_MIN_SPEECH_DURATION_MS
                        Override minimum speech duration in ms
  --vad_min_silence_duration_ms VAD_MIN_SILENCE_DURATION_MS
                        Override minimum silence duration in ms
  --vad_speech_pad_ms VAD_SPEECH_PAD_MS
                        Override speech padding in ms
  --console             Launch interactive Python console for debugging
  --enable_batching     Enable batched inference for faster processing
                        (requires more VRAM)
  --batch_size BATCH_SIZE
                        Batch size for batched inference (auto-detect if not
                        specified)
  --max_batch_size MAX_BATCH_SIZE
                        Maximum batch size to try when auto-detecting
                        (default: 8)
usage: modal_infer.exe [-h] [--non-interactive]

feature-modal 运行脚本

options:
  -h, --help         show this help message and exit
  --non-interactive  跳过 questionary，全部通过参数指定（暂未实现）。

[grider-transwithai]

能不能从v1.6 rebase出来

[Randomless]

试着从main rebase了一下 ，看起来没问题

[grider-transwithai]

更改里面还是有着一段

      # PyInstaller 6.17.0 has a conda hook regression (KeyError: 'depends') in some conda environments
      - pyinstaller==6.16.0

这个在v1.6里面revert回了

      - pyinstaller>=6.0.0

你看看把这个更改去掉?

[Randomless]

已修改

[grider-transwithai]

REPO_REF = "v1.4" 这个指向v1.4是必须的吗，可以不依赖特定tag吗，除了这个之外没什么其他问题了

[Randomless]

不是必须的 改成指向main了