In-place ops deadlock when operands share the same storage

In `InPlaceOp2`, we must acquire a write lock on the lhs tensor's storage and a read lock on the rhs tensor's storage. Because Rust's RwLock is not reentrant, this results in a deadlock when both tensors share the same storage (e.g., are narrowed views into a larger tensor).

https://github.com/huggingface/candle/blob/9ede2041fe36b6ece24c69ce48d19de51c74d166/candle-core/src/custom_op.rs#L361-L364

Start in candle-core/src/custom_op.rs around lines 361-364 and inspect how InPlaceOp2 acquires the lhs write lock and rhs read lock. Reproduce the shared-storage case with narrowed tensor views, then verify that the operation no longer deadlocks while preserving the intended locking behavior.

Written by the indexing model from the issue text.