Add low cpu mem support for LLM instantiation

[arnavgarg1]

Reduces memory overhead when LLMs are loaded onto CPU memory but enabling low_cpu_mem_usage that loads the model using ~1x model size CPU memory. See https://huggingface.co/docs/transformers/main_classes/model#transformers.PreTrainedModel.from_pretrained for more info.

This also enables half-precision loading by default, which seems like a reasonable thing to do. It uses float16 since that always works, unlike bf16 which has hardware restrictions.

[github-actions]

Unit Test Results

  4 files  ±0    4 suites  ±0   51m 41s ⏱️ - 4m 14s
34 tests ±0  29 ✔️ ±0    5 💤 ±0  0 ❌ ±0 
68 runs  ±0  58 ✔️ ±0  10 💤 ±0  0 ❌ ±0 

Results for commit 78b9a9e. ± Comparison against base commit 60f1416.

♻️ This comment has been updated with latest results.