[test] fix: repair CP packed SFT functional setup

[yaoyu-33]

Summary

• Split the CP + packed SFT functional checkpoint setup fix out from PR [test] fix: register internal pytest marker #4359 for faster review.
• Create a tiny pretrain checkpoint first, verify it, then use it as checkpoint.pretrained_checkpoint for the SFT run.
• Keep checkpoint.load = None so this exercises the intended finetune-from-pretrained path instead of trying to resume.

Validation

• git diff --check HEAD~1..HEAD
• uv tool run ruff check tests/functional_tests/test_groups/training/test_seqpacking_cp_example.py
• uv tool run ruff format --check tests/functional_tests/test_groups/training/test_seqpacking_cp_example.py
• uv run --no-sync python -m py_compile tests/functional_tests/test_groups/training/test_seqpacking_cp_example.py
• uv run --no-sync pre-commit run --files tests/functional_tests/test_groups/training/test_seqpacking_cp_example.py

Exact GPU functional execution is left to CI.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

/ok to test dc094b0

[claude]

Review

LGTM. The change correctly splits the test into a two-phase pretrain → finetune flow:

1. Creates a pretrain checkpoint via pretrain() with llama32_1b_pretrain_config
2. Verifies the pretrain checkpoint exists
3. Passes it as cfg.checkpoint.pretrained_checkpoint to the SFT step
4. Sets cfg.checkpoint.load = None so SFT exercises the finetune-from-pretrained path (not resume)

Imports, config fields, and function signatures all check out.

Suggested test cases: No perf tests impacted.