Profile consolidated checkpoint export storage overhead

Tracks the export-storage profiling follow-up identified under NVIDIA-NeMo/Automodel#3576.

Problem

PR #3621 removes repeated full GC scans but does not address possible storage-layer overhead in consolidate_hf_safetensors. Repeated shard lookup, file-open or mmap churn, and buffer copies are hypotheses rather than confirmed bottlenecks.

Scope

• After PR #3621, profile representative consolidated exports.
• Attribute time and memory to shard lookup, open/mmap churn, page faults, tensor-fragment reads and copies, gathering, and conversion.
• If measured, evaluate an FQN-to-source-shard index, bounded file/mmap handle reuse, and zero-copy memory views where lifetime is safe.
• Implement only a measured improvement in a focused follow-up PR.

Acceptance criteria

• Document the representative model, checkpoint layout, storage location, topology, and exact revision.
• Record end-to-end export time, phase timings, host/device memory, page faults, and output size.
• Verify output indexes and tensors against the reference; preserve tied weights, distributed gathering, asynchronous buffer lifetime, bounded resources, and atomic completion.
• Clean up reused handles or mmaps deterministically.
• If the suspected overhead is not material, record the negative result and close without a production change.

Dependency

Use PR #3621 as the baseline and keep this work separate from that PR.