fix pthash memory estimate (#97)

* Account for input hash vector in construction memory estimate

estimate_num_bytes_for_construction returned max(map_bytes, search_bytes),
but the input hash vector lives across both phases. The partitioned
external builder pushes the per-partition hashes (num_keys *
sizeof(hash_type)) into in_memory_partitions and only adds the build
estimate to its running 'bytes' counter, so the bytes < config.ram gate
underestimated true residency by num_keys * sizeof(hash_type) per
partition.

For hash128 (16B) plus the search-phase peak (~25B/key), peak per-
partition residency is ~41B/key while the gate believed ~25B/key, so
config.ram=2GiB actually allowed ~3.2GB resident. Adding the input
hashes outside the max makes config.ram match observed RSS.

* Drop the redundant in-search hashes term and the comment

The previous commit added the input hashes outside the max with the
correct sizeof(hash_type), so the pre-existing num_keys * sizeof(uint64_t)
term inside num_bytes_for_search was double-counting.

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: jermp

include/builders/internal_memory_builder_single_phf.hpp

@@ -211,10 +211,10 @@ struct internal_memory_builder_single_phf {
             num_buckets * sizeof(uint64_t)                                       // pilots
             + num_buckets * sizeof(uint64_t)                                     // buckets
             + (config.minimal ? (table_size - num_keys) * sizeof(uint64_t) : 0)  // free_slots
-            + num_keys * sizeof(uint64_t)                                        // hashes
             + table_size / 8;                                                    // bitmap taken
 
-        return std::max<uint64_t>(num_bytes_for_map, num_bytes_for_search);
+        return std::max<uint64_t>(num_bytes_for_map, num_bytes_for_search) +
+               num_keys * sizeof(typename hasher_type::hash_type);
     }
 
 private: