Add hasCudaSupport() runtime probe + Python has_cuda_support() API

Summary:
This is D6 in the GPU-accelerated H.265 decoding stack for projectaria-tools (plan v11 at ~/gdrive/plans/2026-04-03-gpu-accelerated-h265-decoding-pat-v11.md, tech design at https://docs.google.com/document/d/1TMeLy0TqvAdlYo0IY3Qm4BrzE8i035z9MzYvGl_CzeM/edit by Lou Yang).

Adds the user-facing `projectaria_tools.has_cuda_support()` API specified in the tech design. The Python helper lets users branch on GPU availability without crashing or guessing — typical use is picking a batch size, warning users about CPU fallback, or reporting environment info.

Five files, bottom-up through the layers:

1. `arvr/projects/compression/xprs/xprs.h` — declares `bool hasCudaSupport()` with docstring explaining the runtime semantics.

2. `arvr/projects/compression/xprs/xprsDecApi.cpp` — implementation. Under `XPRS_HAS_NVDEC` it calls `NvCodecContextProvider::getNvCodecContext()` inside a try/catch and returns true iff init succeeds. Under no-NVDEC builds it returns false unconditionally. The first call drives the actual CUDA bring-up; subsequent calls hit the tri-state-atomic cache added in D3 (single atomic load on the fast path, no mutex contention).

3. `arvr/projects/ariane/aria_research_kit/projectaria_tools/core/python/XprsPyBind.h` — exports `has_cuda_support` on the existing xprs submodule with a Python-style docstring.

4. `arvr/projects/ariane/aria_research_kit/projectaria_tools/projectaria_tools/core/__init__.py` — adds `xprs` to the `from . import (...)` block. Without this, the convenience wrapper at the top-level `__init__.py` would raise `AttributeError: module 'projectaria_tools.core' has no attribute 'xprs'` because the `xprs` submodule wouldn't be imported into the `core` namespace. Discovered while testing the end-to-end Python invocation (see test plan item 6).

5. `arvr/projects/ariane/aria_research_kit/projectaria_tools/projectaria_tools/__init__.py` — re-exports as `projectaria_tools.has_cuda_support()` so users get the documented top-level API per the tech design.

Reviewed By: PiotrBrzyski

Differential Revision: D103253729

fbshipit-source-id: ea7320c561efcb504b8de51703fcf5380b2935f4

Author: YLouWashU

core/python/XprsPyBind.h

@@ -143,6 +143,23 @@ void declareXprsDecoding(py::module& module) {
 
 inline void exportXprs(py::module& m) {
   declareXprsDecoding(m);
+
+  // Runtime probe so users can branch on GPU availability without crashing
+  // the import. Returns true iff the wheel was built with NVDEC support AND
+  // the CUDA driver is loadable on the current machine.
+  m.def(
+      "has_cuda_support",
+      &xprs::hasCudaSupport,
+      py::call_guard<py::gil_scoped_release>(),
+      R"DOC(
+Returns True if NVIDIA NVDEC GPU-accelerated H.265 decoding is available
+in this process.
+
+This is a runtime check that combines compile-time NVDEC availability with
+a successful CUDA driver load. The result is cached after the first call.
+On machines without an NVIDIA driver / GPU, returns False; PAT will silently
+fall back to the CPU decoder for all VRS reads.
+)DOC");
 }
 
 } // namespace projectaria::tools::data_provider

projectaria_tools/__init__.py

@@ -14,3 +14,4 @@
 
 # This makes the modules discoverable when doing dir(projectaria_tools)
 from . import core  # noqa
+from .core.xprs import has_cuda_support  # noqa: F401

projectaria_tools/core/__init__.py

@@ -27,4 +27,5 @@
     sophus,
     stream_id,
     vrs,
+    xprs,
 )