qwen3_5_moe_split: rollback BatchKVCache._idx too, not just .offset

BatchKVCache.update_and_fetch writes at [_idx : _idx + new_S] and
returns keys[:_idx]. Stock DFlash's rollback only decrements .offset,
leaving _idx at the post-verify value. Result: stale rejected-draft
keys stay in the cache, and the next verify's attention attends to
them, corrupting the output.

Fix: decrement _idx by the same amount as offset. Now
update_and_fetch overwrites stale entries cleanly.

This matches observed behavior: first speculative verify produces
fine output, second+ verifies produce gibberish due to stale cache
contamination.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/exo/worker/engines/mlx/patches/qwen3_5_moe_split/dflash_split.py

@@ -152,12 +152,19 @@ def make_split_speculative_next(group):  # type: ignore[no-untyped-def]
             )
             n_accepted = int(gathered_n[MOE_RANK].item())
 
-        # 7. Rollback — same n_accepted on both ranks keeps caches consistent
+        # 7. Rollback — same n_accepted on both ranks keeps caches consistent.
+        # For BatchKVCache, both `offset` AND `_idx` must be decremented.
+        # `_idx` is the buffer-length tracker that `update_and_fetch` uses to
+        # decide where to write next and `fetch` uses to slice returned keys.
+        # Stock DFlash only decrements offset, which leaves stale rejected-draft
+        # entries in the cache that corrupt subsequent verifies.
         rollback = verify_len - n_accepted
         if rollback > 0:
             for c in batch.cache:
                 if hasattr(c, "offset"):
                     c.offset -= rollback
+                    if hasattr(c, "_idx"):
+                        c._idx -= rollback
                 elif hasattr(c, "rollback"):
                     c.rollback(n_accepted)