fix: preserve batch-major LAM action ordering

[zjn20030811]

Summary

• restore LAM latent actions with the batch-major layout used by lam_video
• add a dependency-light regression test using unique per-sample/action tokens

Motivation

forward flattens lam_video as (b p) before encoding. The subsequent inverse reshape interpreted that same flattened dimension as (t b), which silently interleaved action tokens across samples whenever both batch size and action length were greater than one. The tensor shape remained valid, so the issue was not visible in shape-only checks.

Validation

• python -m pytest -q --confcutdir=tests tests/test_lam_batch_order.py (2 passed)
• python -m py_compile tests/test_lam_batch_order.py cosmos_predict2/_src/predict2/models/text2world_model_rectified_flow.py
• git diff --check

Fixes #23.