Minor performance improvements

[bbatsov]

Summary

• Fix ProcessedSource force_encoding for frozen string inputs — (+source).force_encoding(...) was creating a mutable copy but discarding the result, so the encoding fix was silently a no-op when a frozen string was passed. Now assigns the result back.
• Avoid intermediate array allocations in HashNode#each_key/#each_value — the old implementation called pairs.map(&:key).each(&block), allocating a full intermediate array on every call. Now yields directly from each_child_node.

Benchmark results

--- HashNode#each_key ---
  each_key (current):  1055691.9 i/s
  each_key (old):       625038.9 i/s - 1.69x slower

--- HashNode#each_value ---
  each_value (current):  1076484.9 i/s
  each_value (old):       591610.7 i/s - 1.82x slower

--- Allocation counts (100 calls) ---
  each_key: old=1301 allocs, new=600 allocs