Upstream collaboration request: integrate multitask fixes validated in production fork

[ahmedelbamby-aast]

Hi maintainers,

We have been running a production-oriented custom fork and would like to align with upstream while reducing divergence.

What we implemented and validated in our fork:

• Task-aware dataset preflight validation for segment/pose/obb label schemas.
• OBB metric compatibility fix so mAP75(B) is consistently reported.
• YOLOv13 multitask task-head config set for detect/segment/pose/obb with safe scales (n/s/l/x).
• Flash backend initialization controls for deterministic fallback vs turing selection, including DDP propagation path.

What we observed:

• Early schema validation significantly improves failure transparency.
• Safe task-head matrix loads reliably in our environment.
• Deterministic backend initialization reduced ambiguity in benchmark backend reporting.

What we request from upstream maintainers:

1. Guidance on preferred scope and ordering for upstreaming (single PR vs split PRs).
2. Feedback on config naming/placement conventions for task heads and scales.
3. Direction on whether Turing flash backend controls should be merged now or behind stricter gating.
4. Any CI/validation expectations we should satisfy before merge.

For convenience, we opened a PR with code-only changes (without Kaggle scripts/artifacts):

• PR: Add YOLOv13 multitask task heads, preflight validation, and Turing flash backend controls #79

If you prefer, we can split this into smaller focused PRs immediately.

Thanks for your work and for reviewing this.