Faster initialization option in DynamicBucketingSampler + various fixes

[pzelasko]

Specifically:

• user may pass precomputed duration_bins into DynamicBucketingSampler to avoid the initial iteration over cuts
• user may override default audio loading duration mismatch tolerance vs manifest metadata with env var LHOTSE_AUDIO_DURATION_MISMATCH_TOLERANCE=<value-in-seconds>
• fix torchaudio version detection
• fault tolerant Shar audio export
• fault tolerant orjson loading
• IterableDatasetWrapper by default won't reset the sampler when __iter__ is called again, unless the sampler finished iteration