Summary

Defensive coding follow-up to issue #2163.

The original repro:

```
File "nemo_rl/algorithms/grpo.py", line 1280, in grpo_train
truncated = repeated_batch["truncated"]
KeyError: 'truncated'
```

was filed against `v0.5.0`. Since then `run_async_nemo_gym_rollout`
in `nemo_rl/experience/rollouts.py` populates `truncated` on the
final batch (mirroring `hit_max_tokens`, parity with the multi-turn
rollout path), so the original `KeyError` no longer reproduces on
`main`. No behaviour change for existing users.

This PR locks in the contract for future rollout backends: when
`grpo.overlong_filtering=true` but the batch lacks a `truncated`
key, raise a clear `RuntimeError` instead of letting it fail
opaquely deep inside the train loop:

```
RuntimeError: grpo.overlong_filtering=true requires the rollout
backend to populate a per-sample `truncated` field on the rollout
batch, but the current batch is missing it. Either disable
overlong_filtering or update the rollout backend to set `truncated`
(see `nemo_rl/experience/rollouts.py` for examples). Refs issue #2163.
```

Why this matters

• Future backends that forget the field hit the loud error instead of
  a stack trace deep in batch handling.
• The error names the offending config option, the missing field, and
  the file containing reference implementations — actionable in one
  read.
• Adds zero overhead on the success path (a single `in` check before
  the existing access).

Test plan

• `ruff check` — clean
• `ruff format` — clean
• `python3 -m py_compile` — clean
• CI `L0`

Refs: #2163