[tintinhamans]

CaptureScreenshot read the whole front buffer into a display-sized surface on every capture and never used it. That was a full GPU readback and a 29 MB surface at 5120x1440, and the capture failed whenever GetFrontBuffer did. It is removed.

The pixel conversion and resize now reuse buffers, and the encoded bytes are moved instead of copied through the cache, upload list and S3 PUT. It also drops an Unlock() on the back buffer, which is never locked and made debug builds assert on every capture.

Probe screenshots went out as soon as they were captured and at full upload speed, and every client did that at the same moment, so the uploads competed with the game traffic. While a multiplayer game runs they now go out one slot at a time: a client waits its slot index times screenshot_slot_offset_ms (2000, clamped 0 to 10000) plus up to 500 ms of random jitter, and curl caps the PUT at screenshot_upload_bytes_per_sec (131072, clamped 16384 to 10485760). The wait and the cap are off once no multiplayer game is running and for the score screen screenshot. Both keys come from the service config, a missing key keeps its default.

Tested: builds x86 Release z_generals.