reject zero-datasize flv tag in FlvReader::Read

[sahvx655-wq]

While reading FLV tags off a playback buffer I traced a stream that came back as one oversized message with the rest of the buffer swallowed. The DataSize in a video/audio tag header is a 3-byte field read straight off the wire in FlvReader::Read; the code then consumes the single VideoTagHeader/AudioTagHeader byte and cuts msg_size - 1 body bytes. When DataSize is 0 that subtraction wraps to 0xFFFFFFFF (msg_size is uint32_t), so cutn() drains everything left in the buffer into one message and desyncs every tag after it. A malicious or corrupt HTTP-FLV source can trigger it with a single 15-byte tag.

Reject DataSize of 0 with EINVAL before any bytes are consumed, since a valid audio/video tag always carries at least the one-byte codec header that the reader immediately reads. The guard sits next to the msg_size read so the underflow can never reach cutn(), and the same fix is applied to both the video and audio readers. Added a regression test that feeds a zero-DataSize tag and checks the reader errors without draining the buffer.

[wwbmmm]

@sahvx655-wq please resolve merge conflicts

[sahvx655-wq]

Done, rebased onto master and pushed. The conflict was only in brpc_rtmp_unittest.cpp: the AVC SPS out-of-bounds fix (#3371) landed a new test in the same spot mine sits, so I kept both tests side by side and gave each its own closing brace. rtmp.cpp merged cleanly with no changes to the guard.

Both flv_reader_rejects_zero_datasize_video_tag and _audio_tag are intact alongside avc_seq_header_sps_without_zero_byte. Should be clean to merge once CI goes green.