Fix mixed bucket key types during metadata cache refresh

[bghira]

Cache loading and dataset initialization converted aspect-ratio keys to floats while discovery appended string keys. This created separate 1.0 and "1.0" buckets and caused empty-bucket warnings during training.

Keep bucket keys as strings across discovery, cache loading, and sampler restore. Merge equivalent keys without losing scheduled sample occurrences, preserve audio/video/caption bucket names, and remove five duplicate float-conversion helpers plus the sampler's mixed-type lookup.

Closes #3206.

Validation:

• Reproduced failing regressions before the fix for cache reload followed by discovery, all four metadata cache loaders, mixed-key assignment, and Webshart metadata preparation.
• 302 tests passed, covering metadata backends, sample occurrence counts, unchanged-layout checkpoint round trips, factory setup, conditioning alignment, and VAE caching:

  .venv/bin/python -m unittest -v -f tests.test_metadata_backend tests.test_webshart_backend tests.test_huggingface_metadata_backend tests.test_sampler tests.test_dataset tests.helpers.metadata.test_video_bucket_strategy tests.helpers.metadata.test_audio_backends tests.helpers.metadata.test_caption_backend tests.test_factory_edge_cases tests.test_conditioning_split_alignment tests.test_vae tests.test_audio_sampler.TestAudioSampler.test_missing_reloaded_metadata_does_not_require_image_crop_coordinates

• All 3 audio sampler tests passed separately: .venv/bin/python -m unittest -v -f tests.test_audio_sampler.
• Black, isort, and git diff --check passed.

The combined sampler/factory/conditioning/audio test order triggers an existing StopIteration failure in TestAudioSampler.test_sampler_bucket_exhaustion; the same failure was reproduced against an unmodified export of main. The audio sampler suite passes in isolation.