Clean up and take ownership of HF/DCP checkpoint storage

Tracks the HF/DCP storage ownership cleanup identified under NVIDIA-NeMo/Automodel#3576.

Problem

components/checkpoint/_backports mixes historical compatibility code with maintained AutoModel-specific storage behavior. Three files are verified unused on current main: default_planner.py, planner_helpers.py, and _version.py (about 697 lines total). The remaining active implementation is blanket-excluded from coverage and type checking.

Scope

• Remove the three verified-unused files in a focused behavior-neutral PR.
• After the current checkpoint stack lands, move maintained implementation out of _backports into a first-class internal package.
• Replace blanket coverage and type-check exclusions with normal validation or narrow documented compatibility exceptions.
• Centralize duplicated safetensors header and dtype handling where ownership becomes clear.

Acceptance criteria

• No import or runtime references remain to removed files, and focused checkpoint tests pass.
• The public checkpoint API and load/save behavior are unchanged.
• Maintained code has explicit ownership without blanket quality-gate exclusions.
• Minimum PyTorch 2.6 compatibility is preserved, with any version-specific shims kept bounded.
• Relocation is split into focused PRs if doing it together would create review noise or conflict with the current stack.

Out of scope

• Replacing the AutoModel reader/writer wholesale with upstream PyTorch.
• Changing performance paths in the current stacked PRs.