Clamp audio samples to valid range to prevent NaN loss during training

[shashank14k]

• I trimmed a large audio into segments, upon loading, one file had maximum sample value >1e29.
• This resulted in a NaN loss during training
• The audio is likely corrupted, and the loss should not equal NaN in most cases, but clamping the values within a range should handle edge cases

[shashank14k]

I could not attach the wav file as github does not support wav format. I could re-encode it to mp4, however rencoding discards such values.