fix: mark encoded frames as the end of their temporal unit

[devopvoid]

Follows up on the discuss-webrtc PSA An update to frame drop handling in VideoEncoder and EncodedImageCallback, which applies to M152 (branch-heads/7977). WebRTC now expects encoders to:

• set EncodedImage::is_end_of_temporal_unit on every output frame (true for encoders without spatial or simulcast layers), and
• report every dropped frame through EncodedImageCallback::OnFrameDropped.

The flag is an unset std::optional<bool> by default. Without it, WebRTC holds what it keeps of a frame longer than needed, for example the captured frame held for corruption detection. The PSA says this may become a hard error later.

Changes

• Hardware encoders (Media Foundation, NVENC, VA-API H.264/H.265/VP9, VideoToolbox H.265): EncoderOutputProcessor, which builds every frame they output, sets the flag. They all produce a single layer.
• Java encoders (VideoEncoderWrapper): the flag is set on each frame, and the frames the wrapper gives up on are now reported through OnFrameDropped instead of disappearing silently:
  • frames the Java encoder produced no output for, found once a later frame comes out;
  • the oldest pending frame when too many are pending;
  • frames that cannot be converted for the Java encoder.
• Media Foundation and VideoToolbox already reported their drops. NVENC and VA-API drop none (Encode() returns an error code instead).

Tests

• New VideoCodecFactoryTests.javaEncoderDroppingFramesKeepsSending: a Java encoder that drops every other frame keeps a call sending, which exercises the new drop reporting. Encoder drops and the flag don't reach getStats in M152, so the test can't check them directly.
• mvn -pl webrtc-jni,webrtc package on windows-x86_64: 285 tests, 0 failures, 3 skipped.
• The Linux and macOS encoders change only through the shared EncoderOutputProcessor; CI covers those builds.