fix(tests): update run_layer_stage callers to the layer_names kwarg

[Anai-Guo]

Problem

run_layer_stage takes a keyword-only layer_names with no default:

# gptqmodel/looper/stage_layer.py
def run_layer_stage(
    looper, *, layers, layer_modules, planning_layer_modules,
    layer_names: Optional[List[str]],   # keyword-only, no default -> required
    fallback, shared_kv_cache_dict, pb, layer_count,
    region_timer, finalize_progress_cls, logger=None,
) -> None:

but three tests in tests/test_stage_modules.py still pass the old
layers_prefix="model.layers", so each dies with a TypeError before reaching
any assertion — it is both an unexpected keyword and a missing required one.

call site	test	binds?
tests/test_stage_modules.py:826	test_run_layer_stage_invokes_subset_stage	TypeError
tests/test_stage_modules.py:1044	test_run_layer_stage_stops_after_last_quantized_layer	binds — already migrated
tests/test_stage_modules.py:1268	test_run_layer_stage_reuses_subset_plan_for_replay	TypeError
tests/test_stage_modules.py:2008	test_run_layer_stage_replays_untouched_layer_outputs_when_all_modules_skipped	TypeError
gptqmodel/looper/module_looper.py:1533	production caller	binds

So one test and the production caller were migrated; three tests were missed.

What the new parameter means

layers_prefix was a single prefix string; layer_names is one full layer
path per layer. run_layer_stage resolves it per index and forwards the
result as create_named_modules' layers_prefix:

layer_name = get_layer_name(layer_names, layer_index)   # stage_layer.py
...
    layers_prefix=layer_name,                           # stage_layer.py

So the fixtures now pass ["model.layers.<i>"] lists sized by layer_count,
matching the already-migrated test at line 1044.

One dummy create_named_modules actually consumes layers_prefix and built
f"{layers_prefix}.{layer_index}.{name}". Since it now receives the full layer
path, the .{layer_index} segment is dropped. The strings it produces are
byte-identical before and after, so the dynamic exclusion pattern that test
exercises (r"-:^model\.layers\.0\.") still matches exactly the same modules:

  layer 0 mlp.gate_proj    old='model.layers.0.mlp.gate_proj'  new='model.layers.0.mlp.gate_proj'  same
  layer 0 mlp.up_proj      old='model.layers.0.mlp.up_proj'    new='model.layers.0.mlp.up_proj'    same
  layer 0 mlp.down_proj    old='model.layers.0.mlp.down_proj'  new='model.layers.0.mlp.down_proj'  same
  layer 1 mlp.gate_proj    old='model.layers.1.mlp.gate_proj'  new='model.layers.1.mlp.gate_proj'  same
  layer 1 mlp.up_proj      old='model.layers.1.mlp.up_proj'    new='model.layers.1.mlp.up_proj'    same
  layer 1 mlp.down_proj    old='model.layers.1.mlp.down_proj'  new='model.layers.1.mlp.down_proj'  same
all full_name strings unchanged: True

The other two dummies accept layers_prefix but never read it, so they are
unaffected.

Verification

The signature was parsed out of stage_layer.py with ast, rebuilt as a stub,
and every run_layer_stage call site in the repo replayed through
inspect.Signature.bind, with the already-migrated test and the production
caller as the control group:

before:  826 -> TypeError: missing a required argument: 'layer_names'
        1044 -> binds
        1268 -> TypeError: missing a required argument: 'layer_names'
        2008 -> TypeError: missing a required argument: 'layer_names'
        3 of 4 test call sites cannot bind

after:   826 -> binds   1044 -> binds   1268 -> binds   2008 -> binds
         0 of 4 test call sites cannot bind
control: gptqmodel/looper/module_looper.py:1533 (production) -> binds

Caveat, stated plainly: I could not execute pytest tests/test_stage_modules.py
to confirm the three tests now pass their assertions. Importing gptqmodel on
this Windows box crashes with a fatal access violation inside
gptqmodel/utils/threadx.py, and the Linux box available to me has no CUDA
build of torch. So this PR is verified at the signature level plus the
string-equivalence argument above; please let CI run it before merging.

Notes

• Test-only change: 4 lines in tests/test_stage_modules.py, no library code.
• The branch is based on my fork's main tip rather than upstream main,
  because my token lacks the workflow scope and both merge-upstream and a
  direct push are refused for .github/workflows/code_quality.yml. The line at
  1044 that upstream has already migrated is deliberately left untouched here
  so the merge keeps upstream's version.

🤖 Generated with Claude Code