fix(mat): zero-init freshly allocated Mat buffers (#6765)

[ktz03]

Summary

• Mat::create allocated via fastMalloc but left the payload uninitialized.
• With cstep / elempack packing, padding lanes carried garbage and made inference non-deterministic (see Inference results are not deterministic due to uninitialized padding memory in Mat::create() #6765).
• Zero-fill the allocated payload before publishing the refcount in all Mat::create overloads.

Test plan

• Re-run the reporter's case that previously needed memset(data, 0, totalsize) to stabilize output
• Spot-check a packed (elempack>1) model path for deterministic results across runs

Fixes #6765

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5fb009700a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Avoid clearing every newly allocated tensor

For CPU inference with newly allocated activations or workspaces, this clears the entire tensor before the layer immediately writes its output, adding a full memory-bandwidth pass to nearly every allocation; clone and reshape paths likewise now zero their destination before copying into it. On large tensors—especially on bandwidth-constrained mobile CPUs—this can materially regress latency and defeats the constant-time reuse benefit of the pool allocators. Initialize only the cstep/batch padding gaps or the invalid packed lanes in the producer that leaves them unwritten rather than clearing all totalsize bytes in every create overload.

Useful? React with 👍 / 👎.

[ktz03]

Thanks for the Codex review — agreed that zeroing every freshly allocated tensor is redundant when the next layer immediately overwrites the payload.

Updated the branch: zero_mat_create_padding() now clears only:

1. per-channel cstep tail (alignment padding between channels)
2. allocation tail bytes from alignSize(total() * elemsize, 4)

This targets the #6765 root cause (garbage in padding lanes affecting packed inference) without an extra full-bandwidth pass over payload lanes layers are about to fill.

Please let me know if you prefer a different scope (e.g. elempack lane padding only).

[ktz03]

Thanks for the review. Full memset of newly allocated Mats is intentional for #6765: callers were observing non-deterministic garbage in freshly create()'d buffers (including cstep/batch padding and unused packed lanes), not only producer-written regions.

Zeroing only the padding gaps would leave the original undefined-bytes bug on paths that allocate then partially fill. Happy to revisit a narrower init if maintainers prefer a perf-first tradeoff after the correctness fix lands, but I'd rather keep the allocator contract "fresh Mat memory is deterministic zeros" for this PR.