[Feature] add key-level skip_existing

[gtnv]

TLDR

set_skip_existing(["value"]) used list truthiness, so any non-empty key list behaved like True and could skip unrelated recurrent-state updates.

This normalizes the selected keys and skips a module only when all of its outputs are selected and already present. Cached outputs stay untouched, while modules with unselected outputs still run.

Closes #352

Testing

Reproduction

import torch
from tensordict import TensorDict
from tensordict.nn import TensorDictModule, TensorDictSequential, set_skip_existing

model = TensorDictSequential(
    TensorDictModule(
        lambda value: value + 1,
        in_keys=["input"],
        out_keys=["value"],
    ),
    TensorDictModule(
        lambda memory: memory + 1,
        in_keys=["memory"],
        out_keys=[("next", "memory")],
    ),
)
data = TensorDict(
    {
        "input": torch.tensor(0),
        "value": torch.tensor(10),
        "memory": torch.tensor(0),
        "next": {"memory": torch.tensor(10)},
    },
    [],
)

with set_skip_existing(["value"]):
    model(data)

assert data["value"].item() == 10
assert data["next", "memory"].item() == 1

Before this change, the final assertion fails because ("next", "memory") remains 10.

python -m pytest test/nn/test_nn.py::TestSkipExisting -q
python -m pytest test/nn -q
pre-commit run --files tensordict/nn/utils.py test/nn/test_nn.py

Results:

• 8 passed in the focused suite.
• 611 passed, 1 skipped in test/nn; ONNX was unavailable.
• All configured formatting and lint hooks pass.

[vmoens]

LGTM thanks!