Support controllable `CutSet.mux` weights in multiprocess dataloading

[pzelasko]

Several changes:

• as in the title, it's possible to use torch's shared memory tensor to supply mux weights and change them in a way that syncs across processes
• simplified DurationBatcher sampling logic (no changes in sampling behavior)
• fixed inconsistency between time constraint exceeded() and close_to_exceeding() (I think it was reported in some issue)
• leveraging dill for CutSet/Sampler inter-process serialization now has to be explicitly enabled with LHOTSE_DILL_ENABLED=1; the library is now less dependent on dill for making CutSet transforms work across main/dataloading processes (you'd only need it if you as the user provide lambdas instead of global functions / partials to map/filter-style functions)

[pzelasko]

Merging, the test that failed is flaky (will have to figure out how to fix it separately, it apparently relies on downloading silero vad...)