Caching doesn't work with cycle

🐛 Describe the bug

When cycle with later caching is used it works the very first time but afterwards it crashes because of the demux because it checks infinitely for todo files but all are cached.

from torchdata.datapipes.iter import IterableWrapper

dp = IterableWrapper(["test"])
dp = dp.cycle()
dp = dp.on_disk_cache(filepath_fn=lambda x: f"./{x}")
dp = dp.map(lambda x: (x, x))
dp = dp.end_caching(mode="t", same_filepath_fn=True)

next(iter(dp))
next(iter(dp))

This happens because of how demux works so the same applies for cycle and demux:

from torchdata.datapipes.iter import IterableWrapper

dp = IterableWrapper(["test"])
dp = dp.cycle()
dp0, dp1 = dp.demux(2, lambda x: 1)

next(iter(dp0))

Versions

torchdata.version=='0.7.0a0+deeacb4'

Start by reproducing the two examples using IterableWrapper with cycle combined with on_disk_cache/end_caching and with demux. Read the cycle, demux, and caching DataPipe implementations to trace the repeated todo-file checks; done means repeated iteration no longer crashes or loops indefinitely when all items are cached.