test: cover converter CLI entry points

[lonexreb]

Summary

Closes the test-coverage gap identified in #2259 — the CLI entry points
(parse_args + main) of the three converter scripts under
examples/converters/ are now exercised by tests, both quickly via
unit tests and end-to-end against real model artifacts in the existing
functional roundtrip test.

The existing tests/functional/test_converter_roundtrip.py only called
the underlying library functions (convert_dcp_to_hf,
export_model_from_megatron, merge_lora_to_hf,
export_lora_adapter_to_hf). The argparse + main() shells of
convert_dcp_to_hf.py, convert_megatron_to_hf.py, and
convert_lora_to_hf.py were never invoked, so regressions in their CLI
contracts (missing flags, broken config-vs-CLI overrides, wrong dispatch
on --adapter-only, etc.) shipped silently.

What's covered

New fast unit tests under tests/unit/converters/

These run without GPU, Ray, or model downloads. The underlying library
calls are mocked out so the tests stay in the L0 budget.

• test_convert_dcp_to_hf_args.py — flags, defaults, unknown-flag
  exit code, plus main()'s local-tokenizer-vs-config-fallback branch
  selection and hf_overrides forwarding (including the YAML
  null → {} collapse).
• test_convert_megatron_to_hf_args.py (marked mcore) — flags,
  defaults, the --no-strict toggle wiring through to the strict=
  kwarg, and the --hf-model-name override path.
• test_convert_lora_to_hf_args.py — required-flag enforcement
  (parametrized over each), --adapter-only dispatch in both
  branches, and the refuse-to-overwrite contract on existing output
  paths.

Extended functional test (tests/functional/test_converter_roundtrip.py)

Added a STEP 10 that invokes each script's main() in-process via
sys.argv patching, against the same model artifacts produced earlier
in the test, and compares the resulting HF checkpoints against the
previously validated state dicts:

• 10a) convert_dcp_to_hf.py CLI on the v1 DCP checkpoint (also
  exercises the local-tokenizer-path branch by saving the tokenizer
  next to the DCP ckpt).
• 10b) convert_megatron_to_hf.py CLI on the Megatron checkpoint.
• 10c) convert_lora_to_hf.py CLI — merged-model branch (no
  --adapter-only).
• 10d) convert_lora_to_hf.py CLI — adapter-only branch
  (--adapter-only), verified by checking the resulting PEFT
  directory layout.

Test plan

• ruff check tests/unit/converters/ tests/functional/test_converter_roundtrip.py — clean
• ruff format — applied
• python3 -m py_compile — clean
• L0_Unit_Tests_Other (CI) — runs the new tests/unit/converters/ suite
• L0_Unit_Tests with --mcore-only (CI) — runs
  test_convert_megatron_to_hf_args.py
• tests/functional/test_converters.sh (CI) — runs the extended
  roundtrip including STEP 10

Notes

• All three new unit-test files have the NVIDIA copyright header per
  repo convention.
• Each main()-exercising unit test uses a fresh importlib-loaded
  module copy so sys.argv patches cannot leak across tests.
• The functional test's new helper _run_script_main invokes main()
  in-process to keep STEP 10 inside the same Ray/CUDA context as the
  rest of the test.

Refs: #2259

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.