CI: g6f instances

[swahtz]

Summary

Moves all three pull-request test workflows onto fractional-GPU (vGPU) EC2 instances and
adds the build option those instances require. Note: A successful run on the g6f instance type run off of this branch can be found here: https://github.com/openvdb/fvdb-core/actions/runs/30422551926

cudaMallocAsync depends on GPU unified memory, which is not available on a vGPU slice —
it fails at runtime with CUDA error 801: operation not supported. NanoVDB selects between
cudaMallocAsync and plain cudaMalloc at compile time via NANOVDB_USE_SYNC_CUDA_MALLOC,
so testing on fractional GPUs requires compiling with that macro defined.

New build option

• FVDB_USE_SYNC_CUDA_MALLOC CMake option (default OFF, so the default build is unchanged)
  in src/cmake/get_nanovdb.cmake, which sets NANOVDB_USE_SYNC_CUDA_MALLOC through
  add_compile_definitions. The macro is consumed in a header by an inline function, so it
  has to be visible to every C++/CUDA translation unit — a per-target definition would give
  the library, bindings, tests, and benchmarks inconsistent inline bodies.
• ./build.sh sync_cuda_malloc build modifier that forwards the CMake define, documented in
  the build-modifier list in README.md.

CI runner configuration

• fgpu_ami added to .github/versions.json, exposed by load-versions.yml as the new
  aws-fgpu-ami and aws-fgpu-az-config outputs, following the existing CPU/GPU pattern of
  fanning the AMI across every subnet for availability.
• tests.yml, cu128.yml, and cu130.yml now start g6f.2xlarge (8 vCPU, 1/4 L4) test
  runners instead of g6.xlarge (4 vCPU, full L4), and build with sync_cuda_malloc.

Restoring async-allocator coverage

Because every PR now tests the synchronous allocator, nightly-publish.yml gains a GPU test
stage on a full g6.xlarge that installs the published nightly wheel — built with the default
asynchronous allocator — and runs the Python unit tests once a night. This is purely additive:
the wheel build matrix is untouched, so no build times or artifacts change. A new
cuda-default-patch output pins the test container to the same CUDA patch version as the
build container rather than hardcoding it.

The nightly stage runs only the Python unit tests, not the gtests, since the gtests already
run on every PR and exercise the same kernels.

Relates to #73, which reports this exact CUDA error 801 failure on an OVX/vGPU VM. The new
sync_cuda_malloc modifier gives affected users a supported way to build for those systems.