[ops] fix: keep verl FusedLinearForPPO batch-invariant under Liger - #48

### What does this PR do?

Makes `fused_kernel_options.impl_backend` authoritative for the PPO
output head:

- `torch` always uses verl's native chunked `FusedLinearForPPOFunction`.
- `liger` explicitly opts into
`LigerFusedLinearScaledCrossEntropyFunction` and retains the native
fallback when Liger is unavailable.
- The selected backend is propagated through all five Hugging Face
fused-forward adapters.

This fixes a regression introduced by #7461. That change selected Liger
whenever it was importable, even when the configured backend was
`torch`. Besides violating the backend contract, the implicit Liger
dispatch bypasses PyTorch operator overrides used by exactness
integrations such as vexact, forcing vexact to modify verl's private
module state as proposed in verl-project/vexact#48.

This does not duplicate an existing PR. Before implementation, open PRs
and issues were searched for `FusedLinearForPPO Liger batch invariant`,
`LigerFusedLinearScaledCrossEntropyFunction`, `fused linear PPO
exactness`, and `disable Liger fused linear`; no matching open work was
found.

AI assistance: OpenAI Codex assisted with implementation and test
execution. The human submitter reviewed every changed line and
understands the change end-to-end. The relevant tests listed below were
run in the shared development workspace before submission.

### Test

- `python -m pytest -q
tests/utils/test_experimental_torch_functional_on_cpu.py` — 6 passed.
- `python -m pytest -q
tests/models/test_fused_kernels_ulysses_sp_on_cpu.py` — 26 passed.
- `pre-commit run --files <all changed files>` — all hooks passed.
- vexact `test_fused_linear_for_ppo_batch_invariance_forward` on 1x H20,
without a vexact monkey patch or environment override — 2 passed.

Additional H20 diagnostic: the full vexact file has two existing strict
bitwise comparison failures between the native chunked 2D matmul path
and the 3D reference matmul, with maximum absolute differences of
`1.52587890625e-05` and `1.9073486328125e-06`. This is separate from the
implicit Liger dispatch fixed here; verl-project/vexact#48 reports 68
passing tests on H100 when using the same native fallback path.

### API and Usage Example

The default remains the native torch backend:

```yaml
actor_rollout_ref:
  model:
    use_fused_kernels: true
    fused_kernel_options:
      impl_backend: torch
```

Select Liger explicitly when desired:

```yaml
actor_rollout_ref:
  model:
    use_fused_kernels: true
    fused_kernel_options:
      impl_backend: liger
```

### Design & Code Changes

- Add an explicit `impl_backend` argument to `FusedLinearForPPO`,
defaulting to `torch`.
- Route to Liger only when `impl_backend="liger"`.
- Store the selected fused backend on each model instance and pass it
through dense, Qwen2-VL, Qwen3-VL, Qwen3.5, and GLM4V torch forward
adapters.
- Document `liger` as an explicit backend and clarify the `torch`
backend behavior.
- Add unit coverage for backend dispatch, validation, and adapter
propagation.

### Checklist Before Submitting

- [x] Read the contribution and agent instructions.
- [x] Search for duplicate issues and PRs.
- [x] Add/update tests and documentation.
- [x] Run pre-commit checks.
- [x] Disclose AI assistance.