[post training] support more data format

[SLR722]

What does this PR do?

Support instruct format (with input and output columns) and chat format (multi turn conversations with conversations / messages column) in post training

We believe this is important for Llama stack post training APIs alpha release and a great enhancement to user experience

Test

1. verified the alpaca dataset generate exact same data before and after this change

2. verified the instruct dataset generate exact same data between torchtune OSS and llama stack, the dataset I used for test https://huggingface.co/datasets/vicgalle/alpaca-gpt4

3. verified the chat dataset generate exact same data between torchtune OSS and llama stack, the dataset I used for test https://huggingface.co/datasets/shibing624/huatuo_medical_qa_sharegpt

[yanxi0830]

Might be helpful to use DataSchemaValidatorMixin for data schema validation https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/utils/common/data_schema_validator.py#L65

[yanxi0830]

Synced offline: we can keep logic for validation in common utils, but there's no value in having them as mixin. Refactor in #720

[yanxi0830]

I think this enum class should be merged with

llama-stack/llama_stack/providers/utils/common/data_schema_validator.py

Lines 19 to 25 in 485476c

	class ColumnName(Enum):
	input_query = "input_query"
	expected_answer = "expected_answer"
	chat_completion_input = "chat_completion_input"
	completion_input = "completion_input"
	generated_answer = "generated_answer"
	context = "context"