native: save() of a saved record merges every property, not the ones that changed #245

save() of a saved record on the native build runs UPDATE :rid MERGE :row with row = every property the object holds, whether or not it changed. Since #194 the stored row is compared with the copy of the values as read before the statement runs, so every value that is as read is already stored: sending it again costs parameter encoding, a larger crossing payload and an engine merge of values that do not change. For a record with a vector (768 floats) and 20 other properties, one changed property sends the whole vector each time.

Proposal (Python only, src/arcadedb/_native/records.py): send only the values that changed, as the caller holds them (an ndarray stays an array). A save that changes nothing still runs its UPDATE with every value, so update events fire as before.

What this does not change: the insert path. save() of a new record adopts the stored row from the statement's reply (the engine converts values to the declared property types, which the record object does not know), so that reply is needed; the update path already skips its read-back for values stored as sent (#194), and command rows are already pulled but not encoded until read (arcadedb_n_cmd). A cheaper insert reply needs a change in the native library (an image rebuild), which is out of scope here.

Measurement: a step-0 probe (untimed) and then a timed A/B on the #194 workload, written to the repo's study standard before any run. The PR is stacked on #194 (the save path it changes) and stays draft until the study is read.