perf: scan adjacent element directly in Next/Prev sibling traversal

[jvoisin]

Next() and Prev() request a single sibling per source node, and distinct source nodes always have distinct immediate siblings, so the results can never contain duplicates. Routing them through mapNodes (which allocates a dedup map and a throwaway one-element slice per node) and the general getChildrenWithSiblingType iterator is suboptimal.

This commit adds a dedicated fast path in getSiblingNodes that scans straight to the adjacent element node and appends into a presized slice, skipping the dedup map and the iterator closure entirely.

benchstat (count=10, all p=0.000):

              |   old    |           new            |
              |  allocs  |  allocs   vs base        |
Next-8        |   60.00  |  2.000    -96.67%        |
Prev-8        |   60.00  |  2.000    -96.67%        |
NextFiltered-8|   65.00  |  7.000    -89.23%        |
PrevFiltered-8|   65.00  |  7.000    -89.23%        |

sec/op:  Next -89%, Prev -92%, Filtered variants -72%
B/op:    ~-80%