Fix ARM half conversion in PackTiles

[gripeyes]

Compatibility:

patch

Issues/Tickets:

Render AOVs from Hou on Silicon macOS contained NaN without this fix, this fix got it working.

Release notes comment:

Fixed scalar half-float packet conversion in PackTiles on Apple Silicon.

Comments for the reviewer:

The previous ARM/NEON implementation used vector load/store intrinsics with scalar addresses, which could corrupt H16 packet values. This change converts through scalar __fp16 values and uses std::memcpy to preserve the exact 16-bit representation.

This was the low-level packet conversion issue identified during an earlier Apple Silicon native AOV investigation:
OpenMoonRay/hdMoonray@c7bbd30...4234f27

The separate hdMoonray render-buffer allocation change from that investigation is intentionally excluded from this PR. The PackTiles correction is independent and should persist regardless of whether the related hdMoonray change is adopted.

Look or scene setup change:

No scene setup changes are required. On Apple Silicon, render outputs previously affected by corrupted H16 packet conversion may now contain the intended values.

Special notes for production:

The changed code path is limited to ARM/NEON builds. The existing non-ARM conversion path is unchanged.

Attention/Reviewers:

AI Assisted Development:

Assisted-by: OpenAI Codex / GPT-5.6

Checklist:

• Documentation has been updated.
• Includes new unit tests.
• Includes new RATS tests.

No documentation, unit-test, or RATS changes are included in this narrowly scoped correction.