[BUG FIX] tuners/lora/eva: fix label mask guard using 'in' instead of hasattr for dict

[Dev-X25874]

What

This is a clean re-open of #3233 (closed due to formatting issues across commits).
All review comments from @BenjaminBossan have been addressed from the
start in this PR.

In prepare_model_inputs_fn_language_modeling, the label mask guard uses
hasattr(model_input, "labels") to check whether labels are present. However,
model_input is a plain Python dict (enforced by the isinstance check three
lines above), and hasattr on a plain dict always returns False regardless of
its keys — dict keys are not exposed as object attributes.

This means the label masking branch (use_label_mask=True) has silently never
executed, even though it is the default in EvaConfig. As a result, ignored /
padding positions (label value −100) were included in the SVD computation
during EVA initialization.

Fix

Replace hasattr(model_input, "labels") with ("labels" in model_input), which
is the correct Python idiom for dict key membership. The body of the branch
already uses model_input["labels"] (subscript access), so only the guard was
wrong. No downstream side-effects for callers that do not include a "labels" key.

# Before
if peft_config.eva_config.use_label_mask and hasattr(model_input, "labels"):

# After
if peft_config.eva_config.use_label_mask and ("labels" in model_input):

Tests

Added test_eva_label_mask, parametrized over use_label_mask=True/False, which
runs get_eva_state_dict on a dataset augmented with a labels column and
asserts the returned state dict is non-empty in both cases. This directly exercises
the previously dead code path.

cc @BenjaminBossan @sirluk

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Okay, looks good from my side. Let's wait a week or so to see if Sirluk has time to check. Otherwise, I'll go ahead and merge.