[Denverjin]

Summary

Allow multiple L1-to-L0A/L0B copies to assemble one GEMM tile, including transposed MajorK/MajorMN loads. Previously, nonzero destination origins were rejected, and each copy derived its compact pitch from its own region.

• Add an Ascend-only T.copy(..., stride=...) argument for the L0 pitch in 16-element M/N fractal blocks. The caller supplies the consuming GEMM's pitch; omitting it preserves compact-region behavior. L0C-to-UB/GM copies also accept an explicit source pitch.
• Support aligned destination offsets, validate fractal alignment and physical write bounds, and map offsets using the copy pitch while preserving allocation and buffer-version spacing.
• Keep BufferRegion access tracking. L0 matrix regions may overlap unless their pitches are provably equal; dependency and lifetime analysis account for fractal padding.

Validation

• Incremental Ascend native build passed.
• 585 host tests passed across copy/GEMM lowering, fractal storage normalization, and AutoSchedule (excluding the two device integration suites).
• 14 NPU tests passed on Ascend 950 (dav-3510), covering partial copies, transposed loads, effective-K tails, dynamic pitches, overwrites, and buffer versions.
• Scoped pre-commit checks and git diff --check passed.

Validation was performed on head 9ec2c59; the merge with the latest main has not been tested.