Add in-memory data support to the JSON DataModule

[Excelius-Wang]

Fixes #1976

Adds an optional json_data input to the supervised-finetuning JSON
DataModule. It accepts either a raw JSON string or a decoded list of
sample dictionaries while preserving the existing file and directory
inputs.

The change:

• requires exactly one of json_path and json_data
• validates in-memory records before dataset splitting
• supports natural jsonargparse CLI input and nested JSON values
• preserves string scalar types during config serialization
• excludes raw training samples from the dataclass representation
• preserves the existing positional constructor arguments

When converting a pandas DataFrame, the supported format is
df.to_json(orient="records").

Tests cover the full train and validation DataLoader path, malformed and
unsupported payloads, CLI parsing and serialization, nested JSON values,
default validation splitting, and backward compatibility.

Testing:

• python -m pytest -q tests/data/test_json.py
• python -m pytest -q tests/data
• python -m pytest -q tests/test_cli.py
• applicable pre-commit hooks
• jsonargparse 4.37.0 compatibility check
• real litgpt finetune_lora --print_config invocation