[Proporsal]Stop injecting the chat template into `messages_log_batch[i]["content"]`; store raw text instead

**TL;DR**

* **Problem:** In GRPO, Environment input `messages_log_batch[i]["content"]`contains **chat-template–applied text(ex: ` {"content": "<|user|>Who are you<|assistant|>"}`), which prevents the Environment from accessing the original prompt.

* I want to use the original prompt for reward modeling, as in the LLM-as-a-judge case.
* **Proposal:**

* Change `content` to store **chat-template–free (raw) text**.
* Keep **`token_ids`** as the chat-templated representation for training/inference .

## Background (current behavior)

* In NeMo/RL’s GRPO pipeline, the [Data Processor](https://github.com/NVIDIA-NeMo/RL/blob/b6269f70f40fd33ca5adfd1f9aa0cac4619581d4/nemo_rl/data/processors.py#L68) **applies the chat template to each message**, stores that templated string in `content`, and generates **`token_ids`** from it to add to the `message_log`.
* Consequently, `messages_log_batch` (the conversation history passed to `Environments`) contains **templated strings** like `<|user|>Who are you<|assistant|>` in `content`.

## Problems

* It is difficult to **faithfully reconstruct the original prompt** from chat-templated text.

* Many use cases require the **original prompt (raw text)** for reward and next-step computation:
* LLM-as-a-Judge (reward from prompt + response)
* Multi-turn environments that generate responses via another LLM
* etc.
* Having the chat template embedded in `content` is unintuitive and harder for humans to read.

## Proposal

Change the Data Processor to behave as follows for each `message_log` entry:

* **`content`**: **raw text before chat template** (the source of truth for humans/Environments)
* **`token_ids`**: token IDs produced **after** applying the chat template (used for training/inference)
* **`content_templated` (optional)**: the **chat-templated text** corresponding to `token_ids` (helpful for debugging/inspection)

### Sample (Before / After)

**Before**

```json
{
"role": "user",
"content": "<|user|>Who are you<|assistant|>",
"token_ids": [ ... ]
}
```

**After (proposed)**

```json
{
"role": "user",
"content": "Who are you?",
"content_templated": "<|user|>Who are you<|assistant|>", // optional
"token_ids": [ ... ] // training/inference unchanged
}
```

## Impact

My understanding is that GRPO (generation/training) **primarily reads `token_ids`** and does not reference `content`, so switching `content` to raw text should **not change training behavior**.

However, **DPO/SFT** paths (e.g., `get_formatted_message_log()`) currently **expect templated text in `content`**, which we should keep in mind.

## Alternative

We could introduce a separate `raw_content` field that holds the chat-template–free string. This minimizes disruption to existing code, but continuing to place templated text in `content` is not common practice. For that reason, I prefer the original proposal (raw in `content`).

## Related issue

* **#682:** I think adopting this proposal would make that issue easier to resolve.