feat(dashboard): preview launcher validation and command

[Dayuxiaoshui]

What does this PR do?

This PR adds a focused preview and validation step to the dashboard train and serve launcher before a subprocess is submitted.

The launcher now:

• validates required model and dataset references, local path existence, numeric ranges, tensor-parallel divisibility, explicit GPU counts, and conflicting options;
• reports field-scoped errors and warnings without loading model weights, initializing workers, or writing launcher artifacts;
• requires explicit acknowledgement for warnings such as unresolved remote ModelScope or Hugging Face references;
• displays both structured resolved arguments and the exact shell-safe command;
• uses the same resolved command builder for preview and submission, preserving argument parity;
• rejects invalid submissions with HTTP 400 and unacknowledged warnings with HTTP 409 before process creation.

The change adds no dependency, external database, hosted service, or public CLI option. Existing train and serve command builders remain available, and the packaged dashboard assets are rebuilt with the new launcher UI.

Related issue

Fixes #269

Issue: #269

Type of change

• Bug fix
• New feature
• Breaking change (public API / CLI behavior changes in a non-backward-compatible way)
• Documentation update
• Refactoring
• Performance improvement
• Test coverage improvement

How was it tested?

Focused CPU and existing train CLI configuration tests:

uv run --no-project --with pytest python -m pytest -q tests/test_dashboard_launcher_cpu.py tests/test_train_cli_config_cpu.py

Result: 91 passed.

Dashboard production build:

npm run --prefix dashboard build

Result: Vite transformed 1,830 modules and completed the production build.

The running dashboard API was also probed with deterministic local references:

• valid preview returned HTTP 200 with matching tokenized and shell-safe commands;
• invalid tensor-parallel topology returned HTTP 400;
• an unacknowledged remote-reference warning returned HTTP 409.

The broader CPU suite was run with:

uv run --no-project --with pytest python -m pytest -q tests/ -k cpu

Result: 401 passed, 5 failed. The five failures are environment-sensitive and outside the changed launcher paths: one writable-path diagnostic expectation, two CUDA-visible-device assumptions, and two distributed NCCL/Gloo subprocess tests that attempted to use the same visible GPU. No GPU training or serving run was performed.

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description (if any).
• Existing tests pass (pytest tests/ -k cpu).
• New behavior is covered by tests.
• Described the test commands run and any hardware limitations.
• Public API / CLI changes are additive and backward-compatible (see CONTRIBUTING.md).

[Dayuxiaoshui]

cc @xsuler