perf(video): 消除中间帧复制并优化 MJPEG 有界预取

通过 Bytes 所有权和只读 AVBufferRef 引用池化 NV12 输入，消除中间 AVFrame 的整帧复制，并保留 V4L2 编码包脱离硬件缓冲的处理。

允许输出健康时有界预取 MJPEG，在输出背压时优先解码前丢帧；增加等帧、输出等待、丢帧原因及编码分段耗时统计，提供预取和所有权输入的对照开关。

补充缓冲寿命、只读保护、软件编码回退和队列策略测试及接口说明，保留现有 PTS、解码 worker、音频和加密行为。

libs/hwcodec/README.md

@@ -41,6 +41,22 @@ Based on the information above, there are several optimizations and changes made
 * remove hevc_vaapi because of possible poor quality
 * amf: not tested, https://github.com/GPUOpen-LibrariesAndSDKs/AMF/issues/378
 
+#### Owned NV12 input
+
+The `bytes` feature provides `Encoder::encode_owned_bytes` for packed NV12
+input to FFmpeg V4L2 M2M encoders. It holds a `Bytes` owner in a read-only
+`AVBufferRef` until FFmpeg releases its last reference, instead of copying the
+input into an intermediate AVFrame. Other encoders retain the existing copy
+path. Converted or stride-compacted One-KVM input also retains that path.
+
+`ONE_KVM_V4L2_OWNED_INPUT=0` disables this optimization when the encoder is
+created, allowing comparisons with the same binary. `ONE_KVM_VIDEO_STATS=1`
+enables preparation, send-frame, receive-packet, callback and output-copy
+timings. One-KVM reports them in `[VideoInputPerf]` at five-second intervals;
+`copied_bytes` counts only the intermediate raw-input copy, not FFmpeg's
+remaining copy into V4L2 MMAP buffers. Encoded packet detachment remains in
+place to avoid retaining scarce V4L2 capture buffers in downstream caches.
+
 ## System requirements
 
 * intel