fix: misc bug fixes and bump fsspec for pyarrow>=17 compat

[cmgzn]

Summary

A collection of small bug fixes and a dependency bump.

Changes

1. fix(image_diffusion_mapper): handle missing or null captions gracefully — prevents crash when samples have missing or empty caption fields. (Related: [Pulp] Handle missing or null captions in image_diffusion_mapper cmgzn/data-juicer#60)

2. fix: warn when process operator list is empty — adds a warning in both DefaultExecutor and RayExecutor when the user provides an empty process operator list, preventing silent no-op execution. (Related: [Pulp] Empty process list produces no log initialization cmgzn/data-juicer#69)

3. fix(vggt_mapper): guard and initialize __dj__meta__ before access — ensures __dj__meta__ is initialized before VggtMapper accesses it, preventing KeyError on samples without prior metadata. (Related: [Pulp] Guard and initialize __dj__meta__ in VggtMapper cmgzn/data-juicer#93)

4. deps: bump fsspec from ==2023.5.0 to >=2026.1.0 — the old pin is incompatible with pyarrow>=17. Root cause: fsspec<=2025.9.0 mirrors ArrowFile.size from the pyarrow stream via @mirror_from (treats it as an attribute), but pyarrow>=17 changed NativeFile.size to a method. This causes TypeError: unsupported operand type(s) for +: 'method' and 'float' in tqdm during HDFS downloads through ArrowFSWrapper. fsspec>=2026.1.0 replaces the mirror with an explicit @property that calls stream.size(), returning the correct int. Verified end-to-end with fsspec==2026.4.0 + pyarrow==24.0.0 on a Docker HDFS setup (load → filter → export).

   Note: datasets==5.0.0 declares fsspec<=2026.4.0, so the effective resolved range is 2026.1.0–2026.4.0. s3fs versions are co-released with matching fsspec versions (e.g. s3fs==2026.4.0 requires fsspec==2026.4.0), so no resolver conflicts.