[LLM] Add Prefix Tuning, PTuning, LoRA, AdaLoRA and Adaption Prompt for LLM fine-tuning

[tgaddair]

This PR adds support for parameter efficient fine-tuning strategies like Prefix Tuning, PTuning, LoRA, AdaLoRA and Adaption Prompt.

To use them in Ludwig with the LLM model types, you need to specify the strategy name in the adapter config:

• Prompt Tuning: prompt_tuning
• Prefix Tuning: prefix_tuning
• P-Tuning: p_tuning
• LoRA: lora
• AdaLoRA: adalora
• Adaption Prompt: adaption_prompt

Here's an example config:

model_type: llm
model_name: facebook/opt-66b
tuner: lora

Things to note:

1. Prompt Tuning, Prefix Tuning and PTuning require the num_virtual_tokens parameter to be set to a non-zero value
2. Lora is only supported for the following model types: t5, mt5, bart, opt, roberta and deberta-v2
3. Adaption Prompt is only supported for Llama model types and requires adapter_len and adapter_layers parameters to be set.

Co-authored by @arnavgarg1

[github-actions]

Unit Test Results

    6 files  ±0    6 suites  ±0   1h 31m 56s ⏱️ + 16m 12s
  33 tests ±0  28 ✔️  - 1    4 💤 ±0  0 ❌ ±0  1 🔥 +1 
100 runs  +1  87 ✔️ ±0  12 💤 ±0  0 ❌ ±0  1 🔥 +1 

For more details on these errors, see this check.

Results for commit 29b6b0c. ± Comparison against base commit e2cde3a.

♻️ This comment has been updated with latest results.

[tgaddair]

LGTM!