Fix local/unfused Megatron attention and LoRA on Ascend

[hazelduan]

Summary

• compare Megatron attention backends through their enum names on Ascend
• keep use_flash_attn disabled and reject padding-free mode for the local and unfused backends
• recognize MCore ColumnParallelLinear and RowParallelLinear when resolving all-linear LoRA targets
• preserve the existing flash/Transformer Engine paths

Root cause

attention_backend is normalized to an enum, but the NPU path compared it with the string "local". The comparison therefore enabled MindSpeed flash attention and generated an NPU flash-attention mask even for local attention. Local MCore attention also does not support padding-free packed sequences.

The same legacy condition treated every backend except local as flash attention. On NPU this silently routed the explicit unfused backend back through MindSpeed FlashAttention. unfused now disables the flash flag and generated flash-attention mask while retaining its distinct backend value.

After selecting a true local MCore layer spec, LoRA target discovery found no modules because it only recognized Transformer Engine linear classes. The local parallel linear classes are now included without changing TE targeting.

Validation

• pre-commit hooks and Python byte-compilation passed for all changed files
• regression coverage covers local padding-free handling and NPU mask selection
• paired mcore-bridge tests cover spec selection, NPU unfused core selection, local norm-weight mapping, and local LoRA dispatch
• A3 validation used SSH user dxq, MindSpeed, Megatron Core 0.16.0, BF16, Qwen3-0.6B, a fixed 500-row dataset, padding_free=false, and identical seed/configuration for local and flash runs
• every successful run completed 100 optimizer steps, kept loss/grad metrics finite, and saved checkpoint-100

Path	Backend	Steps	Mean loss	Last-10 mean loss	Final cumulative s/it	End-to-end wall time
SFT + LoRA	local	100	1.31381	1.04523	0.10189	45 s
SFT + LoRA	flash/TE	100	1.30809	1.04592	0.07201	42 s
SFT + LoRA	unfused	100	1.30841	1.04521	0.10207	54 s
GKD + LoRA, independent frozen teacher	local	100	0.16752	0.15426	0.10990	45 s
GKD + LoRA, independent frozen teacher	flash/TE	100	0.16662	0.15193	0.08711	43 s
GKD + LoRA, independent frozen teacher	unfused	100	0.16602	0.15295	0.10008	45 s

All six valid runs completed 100 optimizer steps. SFT unfused/flash per-step loss correlation was 0.99992 with mean absolute delta 0.00546. GKD unfused/flash total-loss correlation was 0.99912 with mean absolute delta 0.00197; JSD-loss correlation was 0.99419. The GKD runs executed separate student and frozen-teacher forwards on every step and logged total loss, JSD loss, SFT loss, and grad norm.

The A3 image required test-only import shims for unused optional acl/FLA paths; these are not repository changes. A separate Qwen3-30B-A3B teacher attempt reached local MoE construction but exposed an existing unsupported MindSpeedGmmExperts grouped-weight loading layout. The controlled dense student/teacher comparison above keeps the test focused on the attention and LoRA changes in this PR.

Paired PR: modelscope/mcore-bridge#159