Remove triton `AnnAssign` warning filter once no supported torch resolves a triton without the fix (>= 3.8.0)

### What

Many trainer test modules emit a `DeprecationWarning` in CI on Python 3.14 during Triton kernel compilation:

```
triton/compiler/code_generator.py:629: DeprecationWarning: AnnAssign.__init__ missing 1 required positional argument: 'simple'. This will become an error in Python 3.15.
init_node = ast.AnnAssign(target=st_target, value=default_value, annotation=annotation)
```

It shows up across `tests/test_grpo_trainer.py`, `tests/test_kto_trainer.py`, `tests/test_dpo_trainer.py`, `tests/test_rloo_trainer.py`, `tests/test_sft_trainer.py`, and `tests/test_utils.py`.

### Root cause

The warning originates in Triton, not TRL. In `visit_FunctionDef`, Triton builds an `ast.AnnAssign` node without the required `simple` field when a `@triton.jit` kernel argument has both an annotation and a default value. TRL only triggers it indirectly, by compiling Triton kernels during tests; there is nothing actionable on the TRL side.

### Upstream status

Reported upstream: https://github.com/triton-lang/triton/issues/10981

**Not fixed yet** — Triton `main` (3.8.0 dev, latest release v3.7.1) still constructs the node without `simple`. Note this is currently only a `DeprecationWarning`, but it **will become a hard error on Python 3.15**, which the warning filter will not suppress.

### Workaround in place

We suppress the warning via `filterwarnings` in `pyproject.toml`:

```toml
"ignore:AnnAssign.__init__ missing 1 required positional argument:DeprecationWarning",
```

### Action item

- [ ] Remove the `filterwarnings` entry once every supported `torch` resolves a triton containing #10986, confirmed by a warning-free Python 3.14 CI run.
- [ ] **Separate, harder gate: do not conflate with the item above:** before adding Python 3.15 to the CI matrix, triton must actually be upgraded past #10986. On 3.15 the `DeprecationWarning` becomes a hard error, which no `filterwarnings` entry can suppress.

The second item is the reason this issue should not be closed by a filter-removal PR alone: the filter and the 3.15 blocker have independent triggers.

Inspect the filterwarnings entry in pyproject.toml and the affected trainer test modules, then verify that every supported torch resolves a Triton containing the upstream fix and that Python 3.14 CI is warning-free. Done means removing the filter only after that verification; Python 3.15 remains blocked until Triton is upgraded past the fix.

Written by the indexing model from the issue text.