feat(stats): wire real scale/transfer/encode timing from EncState

Oracle step 4 (option A) — give the scale_*, transfer_*, encode_* stats
fields real producers instead of misleading zeros. The fields existed in
FrameTimings and PipelineStats already; producers just weren't passing
non-zero values.

  - avhw.rs: new EncodeStages { scale_us, transfer_us, encode_us } struct.
    EncState::encode_frame (HW VAAPI path) now times the filter graph
    separately from avcodec_send_frame, returning EncodeStages. transfer_us
    is honestly 0 because the HW path never reads back to CPU.
    SwEncState::encode_frame (SW fallback path) returns EncodeStages too;
    there import_and_scale bundles GPU scale + GPU→CPU readback into one
    call, so scale_us includes transfer for SW. Documented inline.

  - state.rs: StreamingEncoder::encode_frame return type bumps from
    Result<()> to Result<EncodeStages>; wlr-screencopy path now feeds
    real per-stage timings into FrameTimings instead of just total_us.

  - state_portal.rs: HW portal path (enc.encode_frame) now extracts
    stages.scale_us / stages.transfer_us / stages.encode_us into
    FrameTimings. Removed the now-unused t_encode_start binding.

Deferred (documented):
  - state_portal.rs SW portal path (line 525) calls import_and_scale +
    enc_thread separately and bypasses SwEncState::encode_frame. To wire
    scale/transfer timing there too, either route through SwEncState or
    thread timing out of import_and_scale. Out of scope for this commit.
  - SW path lumps transfer into scale_us. Splitting requires extending
    import_and_scale's return type — left as a follow-up if operational
    need arises (current default is HW VAAPI).

Oracle audit 2026-06-28 step 4 (option A: integrate, not delete).

All 79 unit tests + 3 integration tests pass. clippy: 0 errors.

src/state_portal.rs

@@ -494,7 +494,6 @@ impl StatePortal {
             }?;
 
             let import_us = t_import_start.elapsed().as_micros() as u64;
-            let t_encode_start = Instant::now();
 
             // 设置帧的显示时间戳（PTS），基于已编码帧序号
             // SAFETY: vaapi_frame is the freshly-imported valid AVFrame returned by
@@ -504,15 +503,17 @@ impl StatePortal {
             }
 
             // 送入编码器完成：缩放 → 回读 → 格式转换 → H.264 编码
-            enc.encode_frame(&vaapi_frame)?;
+            let stages = enc.encode_frame(&vaapi_frame)?;
             let total_us = t_import_start.elapsed().as_micros() as u64;
-            let encode_us = t_encode_start.elapsed().as_micros() as u64;
+            let encode_us = stages.encode_us;
 
             self.frames_encoded += 1;
 
-            // 记录帧计时到管道统计（import + encode 内部各阶段暂不可分离，用 total 覆盖）
+            // 记录帧计时到管道统计（scale 来自 filter graph；transfer 在 HW 路径恒为 0）
             let timings = FrameTimings {
                 import_us,
+                scale_us: stages.scale_us,
+                transfer_us: stages.transfer_us,
                 encode_us,
                 total_us,
                 ..Default::default()