QVAC-25501 fix: support header-only safetensors fitting

[DmitryMalishev]

🎯 What problem does this PR solve?

sd_fit_params rejects header-only safetensors because the shared reader requires tensor payloads to exist. This prevents metadata-only fitting of models with safetensors components.

📝 How does it solve it?

• Scope metadata-only reads to fit initialization and measurement, restoring the previous thread-local mode on every exit.
• Preserve normal loading checks and metadata consistency validation, including offset overflow and invalid offset types.
• Reject non-integer/negative dimensions and bound intermediate shape products before tensor creation, including five-dimensional collapse, zero dimensions and F64/I64 conversion.
• Avoid loading LoRA weights or scale scalars during measurement; adapters loaded after initialization use the same metadata-only scope.
• Initialize the GGML timer before timing a cold-start fit call on Windows.
• Document that a successful fit does not validate weight-file completeness.

🧪 How was it tested?

• Windows Release and Linux Vulkan builds: all 12 CTests passed at 8268f58.
• Reproduced the review finding against 8b9c313: all three supplied malformed headers and F64/I64 overflow cases were accepted. 8268f58 rejects all five; the real FLUX header produces identical metadata for all 1,438 parsed tensors. New regression cases fail against the old reader and pass after the fix. Coverage also includes zero-dimension intermediate overflow, invalid shape types, valid scalar/zero-sized/five-dimensional shapes, strict loading, and SD_FIT_ERROR cleanup.
• Regression coverage: complete/header-only/partial payloads, metadata equivalence, nested scopes, thread isolation, invalid headers and offsets, converted dtypes, shards, cold-start fit, and exception cleanup. LoRA alpha/scale tests verify identical graph structure and preserve real scalar reads outside measurement.
• RTX 5090: the previous engine rejects the FLUX.1 Schnell FP8 header; this revision fits it successfully. Reported module memory matches a full-sized sparse reference with the same header. Normal generation still rejects the header-only file.
• RTX 5090: header-only runtime LoRA with alpha fits successfully and has the same memory report as an adapter with payload bytes.
• Windows and Linux addons build successfully through the temporary vcpkg overlay; 121 Bare tests (467 assertions) and 47 Node tests pass on each host.
• Linux addon C++: 270 passed using the repository's existing CI LeakSanitizer configuration; 3 optional FLUX/Wan model cases lacked assets. RTX 5090 SD2 generation: 25 assertions passed. All 7 ABot integration tests passed (206 assertions), including scene creation and resident/streamed/disk parity.
• Current revision: engine CI passed all build lanes, including CUDA and ROCm; its test-enabled Linux lane passed all 12 CTests. Addon CI passed all nine prebuild targets, C++ suites and eight desktop integration jobs.
• The addon's separate Nx workflow still has one Linux integration job queued for a runner. It has not started or reported a failure; the diffusion-specific workflow above already passed the Linux integration suite. The queued job has not been cancelled or bypassed.
• Current revision: Android, Pixel 9 Pro and iOS, iPhone 17 Pro passed against addon 1266832 / engine 8268f58: 23/23 runner groups on each device, no crashes, 48 non-skipped cases and 31 existing skips per device. Direct header-only fitting is covered by native engine tests. Existing mobile settings skip ABot and several large-model generation cases.
• git diff --check passed.

The review fix and revalidation are complete; this PR is ready for re-review. Merge order: merge this PR after review, publish the merged engine pin in the native dependency registry, then replace the addon overlay with that registry version and validate the final addon revision.