feat: support context parallelism for reward model training on DTensor backend

## Summary

Context parallelism (CP) is not currently supported for reward model (RM) training when using the DTensor backend. Enabling `context_parallel_size > 1` in the DTensor config results in errors, and there is no validation to surface this limitation clearly to users.

## Details

When running reward model training with context parallelism enabled on the DTensor backend, two distinct failures occur depending on the parallelism configuration:

1. **CP + TP**: Setting `context_parallel_size=2` with `tensor_parallel_size=2` fails with:
```
ValueError: Unknown parallel style: local_rowwise
```

2. **CP only**: Setting `context_parallel_size=2` without additional TP fails with:
```
NotImplementedError: Operator aten.log_sigmoid_forward.default does not have a sharding strategy registered
```

The `log_sigmoid` operator used in the RM loss does not have a DTensor sharding strategy registered for CP meshes. A naive workaround of redistributing the rewards tensor to `Replicate` placement produces numerically incorrect results (e.g., accuracy > 1.0), suggesting the fix requires deeper integration with the CP mesh throughout the RM forward pass and loss computation.

## Reproduction

```bash
# Works (no CP):
uv run examples/run_rm.py

# Fails — CP + TP:
uv run examples/run_rm.py \
policy.dtensor_cfg.context_parallel_size=2 \
cluster.gpus_per_node=4 \
policy.make_sequence_length_divisible_by=2 \
policy.dtensor_cfg.tensor_parallel_size=2
# Error: ValueError: Unknown parallel style: local_rowwise

# Fails — CP only:
uv run examples/run_rm.py \
policy.dtensor_cfg.context_parallel_size=2 \
cluster.gpus_per_node=2 \
policy.make_sequence_length_divisible_by=2
# Error: NotImplementedError: Operator aten.log_sigmoid_forward.default does not have a sharding strategy registered
```

## Suggested Action

At minimum, raise a clear error when `context_parallel_size > 1` is set for RM training on the DTensor backend, so users are not silently hitting unsupported configurations. Full CP support for RM would be the ideal resolution.