Native GraphBatch.create_vertices: send float32 vectors in a binary side buffer, not as F-tagged JSON text #222

On the native wheel, GraphBatch.create_vertices sends a float32 vector as number text. NativeGraphBatch._op encodes the op arguments with _enc and calls arcadedb_n_gb(handle, op, json), and codec._default writes a float32 ndarray as {TAG: ["F", value.tolist()]} (TAG is the control character U+0001, the key of every tagged value): Python formats every float and the library parses it back. new_edges and new_edge props carry vectors the same way.

The library already has the binary route for query parameters: arcadedb_n_exec_bin takes a float32 side buffer, codec._default writes ["B", offset, count] into the JSON, and Json.decodeTag reads the floats from the buffer into a float[]. arcadedb_n_gb has no side-buffer variant.

Measurement (untimed, laptop, CI wheels of main 413f54f, native 26.11.1.dev0, one run per cell, so orientation and not a ratio to quote). Shapes are the ones the vector lanes load with: SIFT 128-d in batches of 10,000, Deep 96-d in batches of 10,000 and a 64-d hybrid shape in batches of 5,000. The reference loads the same vertices from the same matrix through insert_columns, which crosses by address.

The text is 1.7 to 5.4 times the float32 bytes, and the create_vertices load is 2.7 to 5.8 times the 2-D reference.

Proposal: add arcadedb_n_gb_bin(handle, op, json, floats, n_floats) next to arcadedb_n_gb. Python sends one-dimensional non-empty float32 ndarrays as B tags in a side buffer when the library has the entry, and keeps the F text otherwise (a library built before the entry, or ARCADEDB_NATIVE_GB_BIN=0). Lists, float64 arrays and rows without a vector keep their text, and the stored float32 values are bit-identical to today. The JVM wheel is not touched. The PR that follows has the tests and the parity check.

Not claimed: the table is a probe, not a timed A/B. A wheel-versus-wheel A/B with the usual repetitions comes after the PR's CI wheel exists.