fix(cml): transpose MatrixReg blocks for MSCT

[Gs-ygc]

Summary

• Transpose each MatrixReg bank/lane block before the LLC reorder path for MSCT.
• Keep the existing data order unchanged for non-transposed MSC stores.

Problem

MSCT already walks scratchpad addresses in transposed order, but each MatrixReg read response was still interpreted as a row-grouped block. The block-local layout was therefore not transposed: for a 64 x 64 e32 matrix, the value at [0, 1] was stored at linear index 1 instead of transposed index 64.

This makes the isolated MLC + MSCT test fail and causes a later Q8_0 matrix result to diverge from NEMU under difftest.

Fix

Interpret the FIFO response as a Matrix_MN x Matrix_MN block of ResultWidth elements. When IsStoreTranspose is set, transpose that block before feeding the existing LLC reorder registers.

Validation

• make comp in XSAI:
  • xiangshan.compile passed
  • xiangshan.test.compile passed
• MinimalMatrixConfig / OpenLLC Apple GSIM with NEMU difftest:
  • MLC + MSCT: mismatch before the fix; Load/StoreT Test passed and GOOD TRAP after the fix
  • MLC + MSC regression: GOOD TRAP
  • Q8_0 M=1024, N=128, K=2048: GOOD TRAP at 2.354 M cycles, 23.308% MFU

Scope

This PR fixes the MSCT output layout only. The independently observed MLCT forward-progress problem is not changed here.

[Wonicon]

Included in #38