fix: resample training images in FlexibleDataManager

[hummat]

Summary

FlexibleDataManagerConfig forces train_num_images_to_sample_from=1 but inherits train_num_times_to_repeat_images=-1 from VanillaDataManagerConfig. In CacheDataloader that combination caches one randomly chosen training image at startup and never resamples another, so any method on this config trains against a single view. Affected methods: geo-neus, geo-volsdf, geo-unisurf.

Closes: #18

Changes

• Default FlexibleDataManagerConfig.train_num_times_to_repeat_images to 0, so a new single image is resampled every iteration instead of caching one forever.
• Document why train_num_images_to_sample_from is pinned to 1 (each batch is one reference image plus its source views in FlexibleDataManager.next_train).
• Turn the CacheDataloader "without resampling" message into a warning, so the same footgun is loud for any other config that sets num_times_to_repeat_images=-1 with a subset of images.
• Add tests/data/test_dataloaders.py covering the resampling behaviour and the corrected config default.

Type of Change

• Bug fix

Testing

• Ran sdf-dev-test (required)
• Added/updated tests
• Tested training manually

Ran on the changed files (full sdf-dev-test blocked locally by the tinycudann GPU build):

• uv run --no-sync ruff format --check + ruff check — pass
• uv run --no-sync pyright sdfstudio/... — 0 errors
• uv run --no-sync pytest tests/data/test_dataloaders.py — 3 passed

Checklist

• My code follows the existing style (ruff format + pyright)
• I have updated documentation if needed