Clarify NVENC live view startup failures

[robertjamesprior]

Summary

CreatePipeline returns the raw GStreamer message when pipeline construction fails. On GPU builds, a failed NVENC/CUDA init surfaces as no element "nvh264enc", which reads like a missing plugin even though the encoder is installed and the real limit is GPU memory.

annotatePipelineError appends a short hint when all three hold:

• the pipeline string asks for nvh264enc
• the error message names nvh264enc
• the error is a missing element/plugin error

Every other pipeline error passes through unchanged.

Test

go test ./pkg/gst -run TestAnnotatePipelineError -count=1

The new test covers both missing-element wordings and one unrelated encoder error. Building the full pkg/gst package needs the GStreamer dev headers, so run it where they are installed.

---

Note

Low Risk
Changes affect error messaging and lock ordering on failed pipeline creation only; successful capture paths are unchanged aside from linking -ldl for the CUDA probe.

Overview
When GPU live-view pipeline creation fails with a missing nvh264enc element/plugin error, CreatePipeline now appends a diagnostic hint instead of returning only the raw GStreamer message. A new gstreamer_cuda_context_probe (dynamic libcuda load, init, device count, create/destroy context) reports whether CUDA is missing, misconfigured, out of memory, or healthy—so operators can tell GPU/CUDA limits from a truly absent plugin.

annotatePipelineError only runs when the pipeline string and error both reference nvh264enc and match known “no element” wordings; all other errors are unchanged. On the failure path, pipelinesLock is released before the probe so a slow CUDA check cannot block other pipelines.

Tests cover NVENC error detection, lock release during probing, and the various nvencFailureDetail message branches.

^{Reviewed by Cursor Bugbot for commit 7ae62fc. Bugbot is set up for automated code reviews on this repo. Configure here.}

[sjmiller609]

The error returned could currently return for a few different reasons. It would be improved if we can surface unambiguous errors, so that we don't show the user an out-of-memory error for a potentially different problem.

Please merge this PR in, then adjust related PRs if needed to handle the more specific errors. Thank you #20