…set`, to detect a pattern where a `DynamicSlice` consumes the output of an `AllGather` with a permuted set of offsets. This pattern is equivalent to a `CollectivePermute` and can be optimized accordingly.The logic is divided into four main sections:1.  **Initial Checks:** Verifies that the `AllGather` is a suitable candidate for this optimization. It ensures the operation is performed across multiple partitions with a single replica and uses flattened IDs (i.e., `use_global_device_ids` is enabled).2.  **Locate DynamicSlice:** Finds the `DynamicSlice` user of the `AllGather`, correctly traversing through any intervening `Reshape` or `Bitcast` operations that do not alter the data.3.  **Extract Offset Specifications:** Derives the memory offset-to-partition ID mappings from both the `AllGather` (the source layout) and the `DynamicSlice` (the destination/permuted layout).4.  **Match Specifications:** Compares the source and destination offset maps to derive the permutation. For each memory offset, it pairs the source partition ID from the `AllGather` with the destination partition ID from the `DynamicSlice`.This CL also introduces a few key data structures to support this logic:*   **`PartitionOffsetSpec`**: Represents the mapping from a memory offset to a partition ID for each replica group. This is used to model the data layout produced by the `AllGather` and the permuted access pattern of the `DynamicSlice`.*   **`PermutationPairs`**: A type alias for a list of `(source_id, destination_id)` pairs, representing a permute operation for CP.PiperOrigin-RevId: 783030292