fix(qemu): use PE-local group matmul dimensions

[jiale-wangOwO]

Summary

• Interpret B.IOS.TSize as Core-level aggregate Shared capacity while keeping B.IOT.TSize PE-local.
• Use LB0=local M, LB1=local N, and LB2=K for single Shared-B and Shared A+B cooperative TMATMUL.
• For Shared A+B, validate Shared A as 4 * LB0 rows and select rows using pe_id * LB0 + local_row.
• Map Local-to-Shared and Shared-to-Local TMOV payloads to fixed per-PE regions inside the aggregate payload.

Validation

• ninja -C build-linx qemu-system-linx64
• python3 -m unittest tests.linxisa.test_v058_pto_contract (23/23 PASS)
• v058_group_mma_fp32_4pe_m128_n32_k64: QEMU == gfrun == independent golden, PASS.
• git diff --check

Known gap

The updated benchmark is an old-toolchain ELF without .note.pto.isa; its diagnostic QEMU run times out and is not claimed as current-ISA golden evidence. The existing Shared-A+B carrier must be regenerated for PE-local LB semantics.

[jiale-wangOwO]

Consolidated into #66. This PR's Shared aggregate / PE-local group-matmul work is folded into the integrated v0.58 tile-semantics PR. Closing to keep a single PR; the tail commits on this branch are superseded by the clean equivalents in #66.