The merge sort can use up to 48KiB of buffers when merging blocks.
That can cause a stack overflow, especially on 64-bit systems when not
building with inlined functions.  This was reported as Issue #462.

row_merge_dup_report(): Allocate buf and offsets from a heap.

row_merge_heap_create(): Allocate space for buf[3] too. Fix bogus
sizeof arithmetics that happened to work, because
sizeof(ulint)==sizeof(void*).

row_merge_blocks(), row_merge_blocks_copy(): Allocate buf[3] from heap.

row_merge_insert_index_tuples(): Allocate buf from graph_heap.

rb://258 approved and tested by Sunny Bains