llama32: parallelize single-row fused RMS norm across two harts

[ChiruGuru99]

Summary

• Repins llama.cpp-et to 26b1011518a33e15050796630e829c315ab6704f.
• Adds a specialized two-hart path for the dominant contiguous [2048,1,1,1] fused RMS-norm/multiply workload.
• Splits reduction and output work across both harts in a minion.
• Exchanges the partial sum and scale through L2 SCP with explicit cache maintenance.
• Leaves the generic fallback unchanged.

Runtime commit:
ChiruGuru99/llama.cpp@26b1011

[AFOliveira]

The previously tested runtime revision (26b101…) passed Llama, shared-runtime regressions and SmolVLM2 manually, but the current head now points to a new revision (1c58f0…) with an additional fused RoPE/cache update and has not been validated. Please disclose the complete runtime commit stack and rerun the exact current head through paired trusted Llama, all shared-runtime regressions, and SmolVLM2. Do not rely on the earlier score for this new head.

[ChiruGuru99]

Revalidated the exact current PR head rather than relying on the earlier 26b101... results.

Outer PR head: aae3e8999bc1fc19497fdd83a9435d6e4e0b7869
Runtime revision: ec53d562050cde3aeafa2f7330216dbfd52e967f
Shared merge base: cc4049d86b14e4ef72f827f3bb767b577f18fbcd

Complete candidate runtime stack:

• 9f3629e3b - Q8_0 matrix-engine dispatch
• ff52f9159 - optimized Q8_0 vecdot
• acce7a2ec - vectorized CONT f32
• 74a6a19d1 - cache-aligned ADD/MUL/SUB
• 37218b4ee - register-resident f16 mul_mat
• 334ec451b - double-buffered f32 matrix-engine mul_mat
• 26b101151 - two-hart RMS norm
• 1c58f0cd9 - fused RoPE/KV cache update
• ec53d5620 - fused residual ADD/RMS/MUL

Exact paired trusted Llama:

• Current main f67c2b2a: 13.4300 tok/s, PPL 15.2119
• Candidate ec53d562: 16.2164 tok/s, PPL 15.2109
• Decode improvement: approximately 20.75 percent

All ten shared-runtime regression models were rerun using the exact candidate runtime and completed with benchmark status pass.

Exact paired SmolVLM2 trusted gate passed:

• Main before: 1,821,160,656 cycles, 3.609s
• Candidate: 1,270,000,888 cycles, 2.692s, PPL 22.2808, accuracy 1.00, no CPU vision fallbacks
• Main after: 1,821,324,439 cycles, 3.608s

Please approve and rerun PR workflows for exact head aae3e899....

[ChiruGuru99]

the first six commits (9f3629e3b through 334ec451b) are inherited runtime-base commits and are listed only for complete reproducibility. The changes contributed by this submission are 26b101151, 1c58f0cd9, and ec53d5620. No authorship is claimed for the inherited commits.