PyTorch release notes split distributed sub-areas into separate worksheets

PyTorch, the open‑source deep learning framework maintained by Meta, has updated its release notes generation pipeline with three targeted improvements. The most significant change concerns the `distributed` category: before this commit, all sub‑area labels like `c10d`, `dtensor`, `fsdp`, `fsdp2`, `checkpoint`, `pipeline`, `torchelastic`, `symm_mem`, `ddp`, `composable`, `rpc`, and `sharded` were collapsed into a single `distributed` worksheet, producing one combined `result_distributed.md` file. Now the script keeps each sub‑area label as its own category, emitting separate worksheets (e.g., `result_distributed_fsdp.md`). This is accomplished by appending `distributed.categories` to the master category list and removing the collapse branch in `category_remapper`.

Additionally, the commit eliminates duplicate labels: `optim` and `optimizer` (which referred to the same area) are now mapped to a single `optim` CategoryGroup, the stale `optimizer` frontend category is removed, and the title heuristic for `[optimizer]` points at `optim`. Similarly, `aot autograd` and `aotdispatcher` are merged under `aotdispatcher`. The result is a cleaner, non‑duplicate `common.categories` list that still covers all previously handled areas. This change gives PyTorch maintainers and contributors more granular release notes, making it easier to track changes per specific subsystem without losing any existing coverage.