[train, data] perf: optimize data fetching latency for full-hetero encoder

[Yangsx-1]

Summary

Optimizes data loading for the encoder in full-hetero DP mode by eliminating redundant data fetching. Previously, the encoder iterated through all microbatches sequentially (including those assigned to other PP ranks), causing unnecessary IO and latency. Now each PP rank only fetches its own assigned microbatches.

Changes

File	Change
loongforge/data/encoder_strided_sampler.py	New module: EncoderStridedSampler (map-style), EncoderStridedIterator (streaming), PrefetchIterator (background prefetch)
loongforge/data/multimodal/dataloader_provider.py	Add build_encoder_energon_iterator for Energon/WebDataset path
loongforge/train/sft/utils.py	Add build_encoder_data_iterator for SFT map-style datasets
loongforge/train/pretrain/pretrain_vlm.py	Wire up encoder iterator construction; add get/set_encoder_data_iterator
loongforge/train/training_utils.py	Refactor train_step to consume from dedicated encoder iterator, removing itertools.tee + sequential skip pattern