[fix]: forward KT activation policy to workers

[yyj6666667]

Summary

• serialize nested kt_activation_policy into ACCELERATE_KT_ACTIVATION_POLICY
• preserve an explicitly supplied worker environment value
• skip forwarding when KT is disabled

Without this, KT workers silently fall back to recompute/recompute even when the launch config requests CPU activation retention. The paired KTransformers parser is in yyj6666667/ktransformers@558e1b84.

Validation

• PYTHONPATH=src python -m pytest -q tests/test_launch.py (4 passed)
• qj5090 Python 3.11 parent/worker subprocess validation: worker resolved cpu=retain, gpu=recompute

[yyj6666667]

CI note:

• quality is failing on 17 pre-existing Ruff errors in src/accelerate/utils/__init__.py; this PR does not modify that file, and the same content is present on sft-v5.
• the test matrix fails before checkout in 0-1 seconds because the required runner group aws-general-8-plus is not configured for this repository.

Targeted validation is green: tests/test_launch.py passes (4 tests), the paired KTransformers policy suite passes (42 tests), and a real two-worker accelerate launch smoke on qj5090 resolved retain/recompute on both ranks.