Remove DictWrapper class from LLM model

[w4nderlust]

Summary

• Remove the DictWrapper class from ludwig/models/llm.py and replace it with a plain Python dict
• DictWrapper existed to prevent input/output features from being registered as nn.Module submodules (which would confuse DeepSpeed). A plain dict achieves the same goal more simply since it is not an nn.Module
• Fix items()[0] indexing (which relied on LudwigFeatureDict.items() returning a list) to use next(iter(...)) compatible with dict_items views

Test plan

• Verify LLM model initialization works correctly
• Verify LLM fine-tuning works (features should not appear as DeepSpeed-managed modules)
• Verify LLM prediction/generation works
• Run existing LLM unit tests

[github-actions]

Test Results

   10 files  ±0     10 suites  ±0   1h 46m 4s ⏱️ - 1m 23s
3 619 tests ±0  3 589 ✅ ±0  30 💤 ±0  0 ❌ ±0 
3 708 runs  ±0  3 665 ✅ ±0  43 💤 ±0  0 ❌ ±0 

Results for commit e2e28eb. ± Comparison against base commit 7aa7353.

♻️ This comment has been updated with latest results.