fix(moq-video): NVENC falls back instead of aborting when no NVIDIA driver

[kixelated]

Summary

On a Linux host with no NVIDIA driver, moq-video's NVENC backend aborts the process instead of falling back to software. cudarc (CudaContext::new) and the NVENC SDK both dlopen their driver libraries lazily and panic! when the library is missing (cudarc's panic_no_lib_found, the SDK's "failed to dlopen the NVIDIA encode library"). Under release's panic = "abort" that panic aborts; in tests (Cargo forces unwind) it shows up as a hard test failure.

So Kind::Auto / Kind::Hardware — the default for moq-cli capture — crashes on any GPU-less Linux box, contrary to the documented "falls back to software (see backend::open)" behavior.

This was surfaced by a Kind::Auto round-trip test in #1802 panicking on the GPU-less CI runner.

Fix

Probe the driver libraries with libloading before calling into cudarc / the SDK, and return Error::Codec if either is absent so backend::open moves on to the next candidate (openh264):

• libcuda.so.1 / libcuda.so (cudarc's CUDA driver), and
• libnvidia-encode.so.1 / libnvidia-encode.so (matches the SDK's own candidate list).

A pre-check (not catch_unwind) is required because release builds use panic = "abort", where catch_unwind can't help. libloading is added as an nvenc-feature-gated optional dep, mirroring the existing cudarc / nvidia-video-codec-sdk gating.

Test plan

• cargo fmt --check clean; non-NVENC cargo build -p moq-video unaffected (macOS).
• New #[cfg(all(target_os = "linux", feature = "nvenc"))] test missing_driver_errors_instead_of_panicking: on a host without the driver, Nvenc::open returns Err rather than panicking. This runs on the GPU-less Linux CI runner (cargo test -p moq-video --all-features), which is exactly where the abort happened before.
• Could not compile-check the cfg(linux) path locally (macOS box; nix/rustup cross-toolchain split blocked a zigbuild). Relying on the Linux CI runner to compile + run the new test.

Note

VAAPI (moq-vaapi, dlopen'd libva) may have the same panic-on-missing-library shape; if so, Kind::Auto would still abort at the VAAPI candidate after NVENC falls back. Worth a follow-up check, but out of scope here.

(Written by Claude)

[kixelated]

Folded into #1819 (commit b0e7682): that PR makes hardware encoders always-on, which is what turns this NVENC panic into a crash on every GPU-less Linux box, so the fallback fix belongs there rather than as a separate PR racing it. Same probe (libcuda + libnvidia-encode via libloading) + the GPU-less test, adapted to the always-on dep layout (libloading as a non-optional target.linux dep). Verified in podman: cargo test -p moq-video --all-features --locked passes with the driver absent.

(Written by Claude)