Feature request: Persist preprocessing across runs (avoid full PDF reprocessing after restarts)

### 🚀 The feature, motivation and pitch

## Summary

On a rented NVIDIA B200 instance, I terminated a training job and later resumed training with the same data and config. The pipeline re-ran the entire PDF preprocessing (~2 hours) before training started again. Request: add a persistent preprocessing cache so restarts and repeated runs can skip unchanged work.

## Problem
- Preprocessing is executed from scratch after interruptions (terminate/stop/restart) even when inputs and parameters are unchanged.
- This causes GPU idle time and higher cost on rented hardware.
- It slows iteration and makes training brittle under preemption or maintenance windows.

## Expected outcome
- Preprocessing artifacts are persisted and reused across runs and restarts when nothing relevant changed.
- Reasonable invalidation when inputs or key parameters differ.
- Improves resilience to interruptions without requiring users to re-run heavy preprocessing.

## Impact
- Lower cost and faster time-to-train on cloud/rented GPUs.
- Smoother restarts and ablations; better utilization of limited GPU windows.

### Alternatives

_No response_

### Additional context

_No response_