[Feature] Support transpose in MX L1-to-L0 staging

[jimmychou0]

Summary

Add transpose support to the public MX L1-to-L0 staging operations pto.mte_l1_l0a_mx and pto.mte_l1_l0b_mx.

Motivation / use case

TileLang FP8 blockscaled GEMM needs this for K-on-row layouts. In examples/ascend/test_gemm_padding_irregular_shapes.py, 4 of the 8 FP8 cases use k_on_row=True; their Final TIR requires transposed data and scale-factor staging from L1 to L0.

Regular pto.mte_l1_l0a/b already expose transpose, but the MX wrappers accept only MX coordinate/stride operands and have no transpose field. TileLang currently rejects the MX path when transpose != 0 to avoid emitting semantically incorrect PTODSL.

PR #1235 (pto.raw_fill_l1) resolves the separate OOB-padding L1 fill blocker. MX transpose remains the blocker for the four K-row FP8 cases.

Proposed API / behavior

• Extend the public PTODSL APIs pto.mte_l1_l0a_mx(...) and pto.mte_l1_l0b_mx(...) with transpose=False.
• Carry the control through public VPTO wrapper ODS, parser/printer, verifier, and pointer normalization.
• Extend wrapper expansion and raw MX load representation so data and E8M0 scale staging use the same transpose semantics.
• Emit the ISA-supported transpose configuration/intrinsic in both VPTO LLVM and CANN900 lowering paths.
• Document the control, supported dtype/layout constraints, and its hardware encoding in the micro-ISA documentation.
• Preserve current behavior for transpose=False. Reject transpose=True with a precise diagnostic if an ISA/backend combination cannot represent it.

Acceptance tests

• PTODSL frontend tests for both MX wrappers, shape-derived and full-control forms, with transpose=True.
• VPTO lit coverage verifies the wrapper-to-raw expansion and emitted LLVM/CANN900 intrinsic/configuration for L0A and L0B.
• Negative tests cover unsupported type/layout combinations.
• TileLang source golden verifies ascend_load_cbuf_to_ca/cb 16-argument MX calls carry the Final TIR transpose flag.
• A5 numerical validation covers all four k_on_row=True FP8 cases in examples/ascend/test_gemm_padding_irregular_shapes.py.

Alternatives considered

Do not silently clear the transpose flag or reinterpret the scale pointer in TileLang: either option changes data/scale layout semantics and would make K-row FP8 results incorrect. Restricting migration to K-column cases leaves half of the FP8 irregular-shape matrix unsupported.

Additional context

Relevant local paths:

• ptodsl/ptodsl/_ops.py (mte_l1_l0a_mx, mte_l1_l0b_mx)
• include/PTO/IR/VPTOOps.td (PTO_MteL1L0aMxOp, PTO_MteL1L0bMxOp)
• lib/PTO/Transforms/VPTOExpandWrapperOps.cpp
• lib/PTO/Transforms/VPTOLLVMEmitter.cpp
• lib/PTO/Transforms/VPTOCANN900LLVMEmitter.cpp

Related: #1235