NanoVDB Python: remove memory-handle ceremony

[swahtz]

Three ergonomics changes to the unreleased NanoVDB Python bindings (everything touched here landed after the last release via #2219, so nothing shipped is broken):

1. GridHandle / DeviceGridHandle are sequences — len(h), h[i] (negative indices, IndexError out of range), and iteration, using the same polymorphic typed-grid dispatch as h.grid(n). grid(n)'s return-None contract is unchanged. An explicit __iter__ is deliberately omitted: __getitem__ + IndexError gives the sequence-iteration protocol, whereas a separate __iter__ path would not carry the def-level keep_alive that parents each grid to its handle (accepted limitation: isinstance(h, collections.abc.Iterable) is False).

2. nanovdb.createNodeManager(grid) returns the typed NodeManager directly — the NodeManagerHandle Python class and its mgr() hop are removed. The lifetime chain is preserved: the C++ handle that owns the node-index buffer is moved to the heap and owned by an nb::capsule, the returned manager is bound reference_internal to that capsule, and the existing keep_alive<0,1> ties the manager to the source grid whose memory leaf(i)/lower(i)/upper(i) point into. The gc.collect() temporary-grid lifetime test keeps its exact shape, minus the .mgr() hop.

3. buffer= parameters removed from io.readGrid(s) / deviceReadGrid(s) and all 13 tools.create* primitive factories, and the zero-method HostBuffer / DeviceBuffer Python classes are unbound. The parameter was unreachable ceremony: neither buffer class had a bound constructor, so the default-constructed prototype (equivalent to omitting the argument in C++) was the only value Python could ever pass, and buffer choice is already expressed by which function or submodule you call (tools vs tools.cuda, readGrid vs deviceReadGrid).

Deliberately out of scope: factories still return GridHandles rather than grids, keeping host/device and single/multi-grid usage uniform.

Relationship to #2232 (injectable CUDA memory resources)

Removing the buffer= vocabulary now aligns the Python surface with the planned C++ retirement of HostBuffer/DeviceBuffer (#2232, Step 3: single-space GridHandle<cuda::Buffer> with copyTo). If allocation control comes to Python later, it should arrive as resource injection in that CCCL cuda::mr-shaped model (the issue defers a SharedResource analog "until the Python-bindings need materializes"), not as buffer prototypes — removing buffer= now avoids ever having two competing allocation vocabularies in Python.

Note for #2225

The pending GPU-bindings PR rebinds a rich DeviceBuffer under nanovdb.cuda — a return-side interop object (__cuda_array_interface__/__dlpack__/device_ptr/recordUse/from_external), never a factory input, so it is a different job than the zero-method class removed here. On rebase, the cuda/PyDeviceBuffer.{cc,h} deletion in this PR should resolve as "keep #2225's new files"; #2225 does not touch PyGridHandle.h, PyTree.cc, PyIO.cc, or PyPrimitives.cc, so the rest rebases cleanly.

Testing

CUDA build (sm_120) with NANOVDB_USE_CUDA=ON, NANOVDB_USE_OPENVDB=ON, BLOSC/ZIP/TBB on:

• TestNanoVDB.py: 172/172 OK on a real GPU, including new host + device sequence-protocol tests (mixed-type merged handles, negative indices, IndexError bounds, iteration lifetime under gc.collect()) and the updated node-manager round-trip / bounds / zero-copy-lifetime tests.
• TestExamples.py: 14/14 OK (load_inspect.py now demonstrates len(handle) / handle[i] / iteration; node_manager.py uses the direct createNodeManager return).

🤖 Generated with Claude Code