Read only needed checkpoint snapshot components during finalization

[bradhilton]

Checkpoint finalization reads each per-block snapshot once for LoRA, three times for optimizer components, and once for step scalars. Each pass currently materializes every tensor. Read only the requested component with safe_open/get_tensor, preserving explicit missing-key errors, tensor lifetime, step ordering and the existing merge, manifest and cleanup paths. There is no retained tensor cache or public API change.

Stacked on #979 (a0c6c767); the production diff is limited to the two reader call sites. A follow-up also changes two inherited test-fixture assignments to setattr so CI type-checks their ModuleType stand-ins; the runtime component is unchanged.

Validation:

• 24 new CPU cases pass on both safetensors 0.7.0 and the exact 0.8.0 wheel used by the frozen native runtime. They compare complete checkpoint bytes across dtypes/ranks/optimizer modes and cover missing/corrupt inputs, error identity, cancellation, abort cleanup, prefix/order behavior and tensor validity after handle closure/file unlink.
• The existing 33 packed-snapshot tests pass. Ruff, formatting and lock checks pass. Pinned ty 0.0.59 passes the changed files; the full local project has the same eight diagnostics as the unchanged base. The initial full prek attempt used the shared environment's older ty 0.0.14 and failed type checks; canonical CI remains required.
• On the two-block public fixture, logical tensor objects requested fall from 100 to 20 and logical tensor bytes from 5,160 to 1,032. These are not physical disk reads or copied bytes; mmap and page-cache behavior matter. Both paths still open each block five times.

A bounded warm-cache CPU reader benchmark using safetensors 0.8.0 measured median wall times of 9.14 → 3.02 ms (16 keys/block, 65,536 elements/key) and 55.65 → 16.30 ms (128 keys/block, 2,048 elements/key). It used two blocks, one Torch CPU thread, and eight alternating-order samples of four passes. CPU time agreed with wall time. This measures header/tensor-view reader work only, excluding tensor-data arithmetic, checkpoint merge/write, physical I/O, comparative RSS and GPU performance. Native background stack samples motivated the change but do not establish interference or speedup.

Canonical CI passes on 5fc36d957: lint/format/type/lock, 1,264 Megatron lightweight tests, and 2,083 unit tests (34 skipped, two deselected). TrainerRank GPU validation completed successfully on two H200s: 95 pytest cases passed (two skipped), followed by the CP2 and TP2 public TrainerRank checks. Sky job 1 reported SUCCEEDED and the workflow completed cluster teardown at 2:29:27 AM MDT. These tests establish compatibility, not native workload performance.

The first head's GPU attempt failed before a training job because two free H200s were unavailable; its launch request was cancelled. Its quality run exposed the two inherited fixture type errors corrected by the follow-up above. Those earlier outcomes remain retained. This PR stays draft and unmerged.