Generation backends disagree on how to report "no OpenAI server": [None] vs []

## Summary

`dp_openai_server_base_urls` is meant to hold one OpenAI-compatible base URL per data-parallel rank. When a backend has no HTTP server to report, the four backends disagree about what to put there: Megatron returns an empty list, the other three return a list of `None`s.

Megatron filters the `None`s out and can then use a plain truth test:

```python
# nemo_rl/models/generation/megatron/megatron_generation.py
if (
not self.dp_openai_server_base_urls # L368 — plain truth test works
and self.cfg["mcore_generation_config"]["expose_http_server"]
):
...
self.dp_openai_server_base_urls = [
url for url in ray.get(url_futures) if url is not None # L374-376
]
```
[megatron_generation.py:368](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/models/generation/megatron/megatron_generation.py#L368),
[:374-376](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/models/generation/megatron/megatron_generation.py#L374-L376)

The other three keep a placeholder entry instead:

| backend | what it returns | where |
|---|---|---|
| vLLM (sync engine) | `[None]` | [vllm_generation.py:529](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/models/generation/vllm/vllm_generation.py#L529) |
| vLLM (deferred load, sync engine) | `[None]` | [vllm_generation.py:548](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/models/generation/vllm/vllm_generation.py#L548) |
| TRT-LLM (no HTTP server) | `[None] * dp_size` | [trtllm_generation.py:266-267](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/models/generation/trtllm/trtllm_generation.py#L266-L267) |
| Dynamo (no token wrapper) | `[None]` | [dynamo_generation.py:220](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/models/generation/dynamo/dynamo_generation.py#L220) |

## Why it is worth changing

The placeholder does not stand for anything. In each of those branches there is no server at all, so there is no per-rank slot to hold open — the list length carries no information. What it does do is make the value lie to a plain truth test: `[None]` is truthy, so "we have URLs" reads as true when there are none.

Callers have to know about that and work around it:

- [`_shard_base_urls`](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/algorithms/single_controller_utils/setup.py#L768-L773) cannot write `if not urls`; it needs `if not any(urls)`.
- [grpo.py:1543](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/nemo_rl/algorithms/grpo.py#L1543) prints the count, so a sync run reports `Reserved 1 vLLM server URLs: [None]`.

It also forces the type to be wider than the values ever are. Every one of these lists is either all real URLs or all `None`s — never mixed. If the "none available" case returned `[]`, the type could be `list[str]` instead of `list[Optional[str]]`, and callers downstream would not have to carry the `Optional` around. This came up while reviewing #3367, where a function signature was widened to `list[Optional[str]]` to match the producer's declared type even though a `None` cannot reach that code path.

## Suggested change

Make the three backends match Megatron: return `[]` when there is no HTTP server to report, and narrow `dp_openai_server_base_urls` to `list[str]`.

Worth checking while doing it:

- [`tests/unit/models/generation/test_dynamo_generation.py:167`](https://github.com/NVIDIA-NeMo/RL/blob/e9a855ff4bc0a41967e9491c41de093527e123b6/tests/unit/models/generation/test_dynamo_generation.py#L167) asserts `== [None]` and would need updating.
- TRT-LLM's version is length `dp_size` rather than 1, so anything that reads the length in the no-server case should be checked before changing it.
- `_shard_base_urls` and the `any(...)` guards can be simplified once the sentinel is gone, but they stay correct either way, so they can be cleaned up separately.

Start with the cited no-server branches in vllm_generation.py, trtllm_generation.py, and dynamo_generation.py, then inspect dp_openai_server_base_urls callers such as _shard_base_urls and grpo.py. Update the affected tests, including tests/unit/models/generation/test_dynamo_generation.py:167, and check TRT-LLM length handling. Done means no-server cases return [] consistently and the type and callers no longer require Optional URLs.

Written by the indexing model from the issue text.