Add /stats and /latest_frame endpoints

[alexnorell]

Summary

• Adds GET /stats endpoint that returns aggregated camera_fps, inference_fps, and stream_count across all active pipelines. Reuses existing list_pipelines/get_status IPC, no new commands needed.
• Adds GET /inference_pipelines/{pipeline_id}/latest_frame endpoint that returns the most recent frame as base64-encoded JPEG with metadata (frame_id, frame_timestamp, source_id). Uses a new LATEST_FRAME IPC command that peeks at the buffer non-destructively.

Both endpoints are gated behind ENABLE_STREAM_API.

Test plan

• pytest tests/inference/unit_tests/core/interfaces/stream_manager/test_stats_and_latest_frame.py -v
• Manual: start server with ENABLE_STREAM_API=true, call GET /stats and verify JSON shape
• Manual: start a pipeline, call GET /inference_pipelines/{id}/latest_frame and verify base64 JPEG decodes

[dkosowski87]

description + response_model please ;)

[dkosowski87]

tasks = [
    self.stream_manager_client.get_status(pid)
    for pid in pipeline_ids
]

responses = await asyncio.gather(*tasks, return_exceptions=True)

? Not sure how much would it matter though.

[dkosowski87]

Couldn't we split this into somthing like :

async def get_stats():
  reports = await self.fetch_pipeline_reports()
  compute_stats(reports)

I think we are pushing too much into the endpoint function bodies. This is not a place for business logic.
Additionally thinking to write in this way will allow use to easily optimize the work. I know that the stats calculation is trivial, but if it would be more complicated - this is a blocking operation. Having it as a separate compute_stats would allow us to quickly fix this, running this in a separate thread or something like that.

[dkosowski87]

Wouldn't we like to allow to parametrize this through the request? Not sure about it's usefulness at this moment because I don't know the full context, but just wanted to point this out. Although in that case I would provide some enum with some reasonable values, low, medium, high where medium is 70 for example. Otherwise people would probably skew to typing 100 all the time.