Fix loop variable shadowing in validation that corrupts sample iteration

[Mr-Neutr0n]

Summary

• Fixed a variable shadowing bug in Trainer.validate() where the inner for i, ... in enumerate(validation_artifacts) loop shadows the outer for i in range(num_validation_samples) loop variable.
• After the inner loop completes, i retains the last enumerate index rather than the outer sample index, which can cause incorrect validation behavior with multiple samples.
• Renamed the inner loop variable from i to artifact_idx to eliminate the shadowing.

Details

In finetune/trainer.py, the validate() method has a loop over validation samples:

for i in range(num_validation_samples):       # outer loop (line 540)
    prompt = self.state.validation_prompts[i]  # uses i
    image = self.state.validation_images[i]    # uses i
    video = self.state.validation_videos[i]    # uses i
    ...
    for i, (artifact_type, artifact_value) in enumerate(validation_artifacts):  # shadows i! (line 591)
        artifacts.update({f"artifact_{i}": ...})

In Python, for loop variables are scoped to the enclosing function, not the loop body. The inner for i overwrites the outer i. While the outer for i in range() reassigns i at the top of each iteration, this shadowing is a correctness hazard that can produce subtle bugs if the code is modified, and violates the principle of least surprise.

Test plan

• Verified only the inner loop variable name changes via git diff
• Run validation with multiple samples to confirm correct indexing
• Run with DeepSpeed ZeRO stage != 3 to confirm process distribution works