[CICD] Add train image pipeline and real TE-FL smoke testCi/ascend train te smoke

[AlexMa616]

Description

Add an Ascend training image pipeline and enable the Ascend CI suite with a
real TransformerEngine-FL training smoke test.

The change builds the training runtime from Ascend-specific Docker and installer
files, validates Megatron-LM-FL and TE-FL NPU initialization, and runs a
two-device Qwen3 0.6B smoke test through the TE-FL reference backend.

Chip-specific setup remains outside reusable common workflows. Ascend behavior
is defined through the platform configuration, Ascend installer and launcher,
Dockerfile, and dedicated image-build workflow. No FlagScale core source is
changed.

Prerequisite: TransformerEngine-FL must include 771a7488
(fix(ascend): include NPU backend in Python package) or an equivalent upstream
change before rebuilding the image from the official TE-FL repository.

Type of change

• Infra/Build change (changes to CI/CD workflows or build scripts)
• Bug fix
• Code refactoring
• New feature (non-breaking change which adds functionality)
• Documentation change
• Breaking change

Changes

• Added a trusted Ascend training image build and Harbor publishing workflow.
• Added the Ascend training Dockerfile and source dependency installation for
  Megatron-LM-FL and TransformerEngine-FL.
• Added an Ascend launcher that restores TE-FL NPU compatibility after Megatron
  platform initialization.
• Added a two-device Qwen3 0.6B mock-data smoke test using
  transformer_impl: transformer_engine and TE-FL reference.torch operators.
• Added platform-neutral finite-loss smoke validation and parser tests.
• Enabled Ascend through the standard all_tests workflow and selected the
  validated Harbor training image.
• Aligned torchvision with the validated torch 2.7.1 and
  torch_npu 2.7.1.post2 runtime.
• Improved installer cache path handling, resumable sccache downloads, and
  relative requirements include processing.

Validation

• Repository-built training image smoke:
  • Completed 3 training iterations on two Ascend NPUs.
  • Dispatched TE-FL GEMM, RMSNorm forward/backward, and optimizer operations
    through reference.torch.
  • No skipped or NaN iterations.
  • 1 passed
  • FLAGSCALE_ASCEND_REPO_IMAGE=PASS
• Validated Harbor image:
  • harbor.baai.ac.cn/flagos-dev/flagscale:manual-20260724-ascend-train-426ad116-te771a7488
  • Digest: sha256:2fe15cf8b339b6828395fc3ce56e67b004b7c5f251d9e6673f20579819fe2c16
• GitHub Actions:
  • https://github.com/BrianPei/FlagScale/actions/runs/30083843003
  • CLI, unit, train, inference, serve, and aggregate jobs passed.
  • Benchmark and hetero-train remained skipped because they are not enabled
    in the Ascend matrix.
• Local focused checks:
  • Shell, Python, YAML, and Git diff validation passed.
  • Ascend training matrix parsing passed.
  • Result parser tests: 2 passed.

Checklist

• I have read and followed the contributing guidelines
• The functionality is complete for the Ascend training pipeline and smoke scope
• I have commented my code, particularly in CI workflow setup steps
• My changes generate no new warnings
• I have tested the training workflow on the Ascend platform

[CLAassistant]

All committers have signed the CLA.