feat: add ICONS selector — Influence Consensus for Vision-Language data selection

[tongruiliu]

Summary

Adds ICONS (arXiv:2501.00654) as a new online dynamic_select selector. ICONS scores every training sample against K task validation sets via gradient influence, then aggregates the K per-task rankings by majority vote to pick a subset that is robust across tasks rather than tuned to a single validation target.

What's included

File	Change
src/dataflex/train/selector/icons_selector.py	New IconsSelector: projected-gradient collection, per-task influence ranking, majority-vote consensus
src/dataflex/train/selector/__init__.py	Register icons in the selector registry
src/dataflex/configs/components.yaml	New icons component
examples/train_lora/selectors/icons.yaml	Runnable LoRA example config

Usage

FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 \
  dataflex-cli train examples/train_lora/selectors/icons.yaml