feat: [cosmos3-action] Add example processing script and assets for the egocentric hand action data

[qianlim]

Summary

Addressing the issue #31 at cosmos-framework repo; making the PR in this repo as this is tutorial/example-like code, better suited here.

• Add an egocentric hand-action data processing example for Cosmos3 Action.
• Include a small ESCALE_000374 sample with video, camera, caption, and 3D hand-pose annotations.
• Add a converter script that produces the raw 57D hand-action layout:
  [camera, right_wrist, right_fingertips, left_wrist, left_fingertips].
• Document the expected input JSON schema, coordinate conventions, wrist-frame alignment hook, output files, and optional model-space normalization.

Tests

• Ran the converter from the cosmos repo root using public cosmos-framework imports:

  PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=/home/qianlim/projects/cosmos-framework \
    python cookbooks/cosmos3/generator/action/finetune/data_processing_for_egocentric_hand_action.py \
    --output-dir /tmp/cosmos_pr_readme_verify_final

• Verified output shape: (121, 57).
• Verified roundtrip fingertip reconstruction error: right max/mean: 4.217e-05 / 1.973e-05 m; left max/mean: 3.102e-05 / 1.537e-05 m.
• Ran py_compile on the new script.
• Generated a local visualization overlay from the decoded action for manual sanity check.