Add Transformer Engine compatibility shim, fallbacks, and modernize TE pins

[LilBadger]

Motivation

• Ensure the repo runs on Blackwell GPUs (CUDA 12.8 / cuDNN 9.x) without crashing when private Transformer Engine (TE) internals have moved or been removed.
• Provide stable, public-API-first replacements for TE features (RMSNorm, RoPE, attention helpers) and fall back to pure-PyTorch implementations when TE is absent.
• Relax strict TE version pins so modern TE releases (e.g. 2.11.x and later minor versions) can be used without API breakage.

Description

• Added a compatibility layer at cosmos_predict1/utils/te_compat.py that exposes RMSNorm, apply_rotary_pos_emb, check_set_window_size, split_along_dim, DotProductAttention (if available), and InferenceParams fallbacks; it tries public TE APIs first and falls back to pure-PyTorch implementations when needed.
• Removed reliance on TE private symbol _RMSNorm by switching to the compat RMSNorm in cosmos_predict1/autoregressive/utils/parallel.py and using a safe forward path when TE internals are not present.
• Updated cosmos_predict1/diffusion/module/attention.py to use the compat RMSNorm and the existing compat apply_rotary_pos_emb so modules no longer import TE internals directly.
• Relaxed TE pins in requirements.txt to transformer-engine>=2.11,<3 and transformer-engine-torch>=2.11,<3, and updated scripts/install_blackwell_wsl.sh accordingly.
• Added a smoke-check script scripts/check_te_compat.py that imports main inference entrypoints and prints TE_AVAILABLE, TE_DotProductAttention_available, and whether the TE fast-path is enabled (script returns WARN for environment/system-lib import issues instead of failing).

Testing

• Ran the smoke script python scripts/check_te_compat.py; the script executed and reported TE_AVAILABLE=False, TE_DotProductAttention_available=False in this environment and returned compat_smoke=WARN (imports warned due to missing system lib libGL.so.1 in the minimal CI container, not TE symbol breakage).
• Compiled affected modules with python -m compileall cosmos_predict1/utils/te_compat.py cosmos_predict1/autoregressive/utils/parallel.py cosmos_predict1/diffusion/module/attention.py scripts/check_te_compat.py and compilation succeeded.
• Performed lightweight runtime check from cosmos_predict1.utils import te_compat; te_compat.RMSNorm(16) which returned the fallback RMSNorm type successfully.

---

Codex Task