NanoVDB: make the nano components warning-clean under OPENVDB_CXX_STRICT and enable it in NanoVDB CI

[harrism]

Context

#2280 made every CUDA translation unit in the NanoVDB build warnings-as-errors by default (NANOVDB_CUDA_WERROR, covering both the NVCC front end and the forwarded host-compiler pass). The remaining gap is pure host TUs — .cc sources such as TestNanoVDB.cc, the tools, and the python bindings — which build with no -Werror at all.

The repository already has the mechanism: OPENVDB_CXX_STRICT (-Werror -Wall -Wextra plus the curated per-compiler set in OpenVDBCXX.cmake), which the core library builds with in CI (build.yml). The NanoVDB components have never built under it, and today they cannot: replicating the strict GNU and Clang flag sets over a standalone NanoVDB build (unit tests + tools + examples, Release and Debug) fails on pre-existing warnings.

Measured inventory

gcc (13.3): 74 unique sites

Category	Sites	Notes
-Wunused-parameter	39	mostly named-but-unused params in header interfaces, e.g. the dummy parameter of DeviceBuffer::create
-Wunused-function	17	file-static functions in the portable C headers (CNanoVDB.h, PNanoVDB.h) — see policy question below
-Wsign-compare	10	Debug-only: inside assert bodies, so Release builds never see them
-Wdangling-else	3
-Wpedantic / -Wformat / -Wunused-variable	5	includes %p passed non-void* in an example

clang: 33 unique sites — overlapping categories plus -Wconversion (in the Clang strict set but not the GNU one) in NanoVDB.h, PNanoVDB.h, math/Math.h and tools/CreateNanoGrid.h.

Raw-error counts are much higher (~180 sign-compare instances in a Debug build) because header sites repeat per TU; the unique-site counts above are the real work.

Policy decisions needed

1. CNanoVDB.h / PNanoVDB.h: file-static functions are the design of these portable C-style headers, and PNanoVDB.h is synced with external consumers. Options: attribute/edit every site, suppress -Wunused-function at the including test TUs, or exclude these headers' test TUs from strict. Editing PNanoVDB.h piecemeal seems wrong; a deliberate choice is needed.
2. clang -Wconversion in public headers is semantic-adjacent (narrowing arithmetic in core math paths) and needs careful review rather than mechanical casts.

Suggested direction

Fix in slices, each independently reviewable: (1) the mechanical classes (unused parameters/variables, dangling-else, format); (2) the Debug-only sign-compares in asserts; (3) the C-header policy decision and its implementation; (4) -Wconversion last. Once clean, add -DOPENVDB_CXX_STRICT=ON to the linux-nanovdb matrix in nanovdb.yml (all four gcc/clang × Release/Debug configs), mirroring what build.yml already does for the core. OPENVDB_CXX_STRICT itself should stay opt-in in CMake — it is a repo-wide option — but on in CI; the CUDA-side NANOVDB_CUDA_WERROR remains default-ON as landed in #2280.