[Feature] Add inplace=True to gather / repeat / repeat_interleave / roll / reshape / flatten / unflatten / contiguous

[vmoens]

Summary

Follow-up to #1706 (which added inplace=True to pad). Extends the same leaf-by-leaf-rebind pattern to the rest of the shape-changing ops that were flagged in that PR's punch list:

• Always-copy ops: gather, repeat, repeat_interleave, roll.
• Sometimes-view ops: reshape, flatten, unflatten, contiguous.

Each op gains an inplace: bool = False kwarg. When True, the tensordict's object identity and key set are preserved, leaves are replaced one at a time inside self, and the old leaf storages are released as their replacements are written — so the common td = td.op(...) pattern drops from ~2x peak memory to ~1x.

A new shared helper TensorDictBase._inplace_rebind_leaves encapsulates the iteration / rebind / _change_batch_size sequence so each op only provides the per-leaf transform, the nested-TD recursion, and the new batch_size.

Design notes

• LazyStackedTensorDict raises NotImplementedError on inplace=True for these ops; its overrides reshape/flatten/unflatten/repeat into the stack structure itself, which can't be safely rebound in place. The error message suggests .to_tensordict() as the escape hatch.
• View branch (reshape, flatten): Tensor.<op> may return a view when memory layout permits. The rebind still happens (object identity + batch_size are correctly updated), but the new leaf shares storage with the original so the memory benefit does not materialize. The docstrings call this out.
• contiguous(inplace=True): leaves that are already contiguous return self from Tensor.contiguous(), so the rebind is a no-op for those keys. Useful when the leaves are a mix of contiguous and non-contiguous tensors.
• gather(inplace=True): requires index.ndim == self.batch_dims so the result keeps the same number of batch dims. Mutually exclusive with out=.
• repeat_interleave(inplace=True): requires an explicit dim and ndim > 0 because the dim=None / scalar paths internally reshape, which would change ndim.
• No safe= kwarg here (unlike pad). For these ops there's no realistic per-leaf pre-flight check that doesn't require a CUDA sync — the only mid-loop failure mode is OOM, which can't be pre-validated.

Test plan

• `pytest test/test_tensordict.py -k "repeat or roll or gather or reshape or flatten or unflatten or contiguous or inplace"` — 672 passed, 26 unrelated skips
• `pytest test/test_tensordict.py::TestGeneric -k inplace` — 33 new tests
• `pytest test/test_tensorclass.py` — 144 passed (stub tests pass with new `inplace` kwargs added to `tensorclass.pyi`)
• New tests cover, per op: `inplace=True` returns `self`; leaf-for-leaf equivalence with the `inplace=False` path; `LazyStackedTensorDict` raises `NotImplementedError`; for `repeat` and `roll`, a weakref-based check that old leaf storages are released. `gather` also checks the `out=` + `inplace=True` conflict and the ndim-mismatch error.