test(rtp_recv): add end-to-end smoke test exercising threads + GL

[osamu620]

Background

open_htj2k_rtp_recv --smoke-test currently runs four unit-style
tests against individual components (UDP socket, RFC 9828 parser,
YCbCr→RGB helper, frame reassembler). None of them touch:

• The three-thread pipeline (receive / decode / render)
• The RTP-timestamp pacer
• The GL 3.3 core shader programs
• The AVX2 YCbCr path's end-to-end integration
• Shutdown sequencing (stop_flag propagation, thread join order)

Regressions in any of those surfaces are only caught by running the
binary against a live sender — which isn't practical in CI.

Proposal

Add a self-contained end-to-end smoke test that:

1. Forks (or spawns via std::thread) the existing Python loopback
   helper — or reimplements the minimal RFC 9828 sender inline to
   avoid the Python dependency in CI.
2. Sends a handful of frames (e.g. 10) of a small known codestream
   (conformance_data/ds0_ht_01_b11.j2k) into the threaded receiver
   with --threading on, --frames 10, --no-render.
3. Verifies: exit status 0, frames_decoded == 10, frames_failed == 0,
   zero decode-slot / render-slot evictions, non-empty decode-time
   stats.
4. Repeats with --color-path cpu to cover the AVX2 path end-to-end.

A headless CI runner won't have GL, so --no-render is the correct
default for this test. A separate --render smoke for GL paths can
be added later, gated on a display environment variable.

Acceptance

• New smoke test is wired into --smoke-test or a new --e2e-test
• Runs to completion in <30 s without external network
• Works in a headless environment
• Flags: both --color-path shader (if GL available) and --color-path cpu