torchcodec + H3 single file fixes

[bghira]

This pull request introduces several improvements and bug fixes across audio loading, model configuration, and TorchInductor training pass handling, along with new and enhanced tests to ensure the reliability of these changes.

Audio Loading Improvements:

• The load_audio function now attempts to use ffmpeg for all filesystem paths if torchaudio fails, regardless of file extension, improving robustness when TorchCodec is unavailable. Logging for ffmpeg fallback failures is also improved.
• Added a new test AudioLoadTests.test_filesystem_path_uses_ffmpeg_when_torchaudio_fails to verify that ffmpeg is used as a fallback when torchaudio.load fails.

TorchInductor Training Pass Handling:

• Updated the dynamic training pass configuration to import the TorchInductor config module more robustly and handle cases where the disabled_passes attribute is missing, with appropriate logging.
• Added a test to ensure that missing disabled_passes in TorchInductor config is handled gracefully and the environment variable is still set.

Model Configuration Logic:

• Improved _model_config_path in the MiniMaxH3 model to correctly handle cases where only the transformer path is a single file, ensuring local components are preserved.
• Added a test to verify that _model_config_path returns the correct path when the transformer is a single file.