Do not save or read back resident zero pages; keep zero-split areas on the parallel readers #3

Checkpoint images save resident pages that hold only zeros, and restart reads them back. On the Qwen3.8-Flash-Next-NVFP4 H200 worker, one 47.68 GiB anonymous payload (the FP8 PLE n-gram embedding, pinned on the CPU) is most of the 53.5 GiB worker image and 16.2 s of its 20.9 s cold memory pass on Runpod NFS (release-cold-cycle run 09: read_kib=50000896 read_us=16241444 mode=pread). A 15 GiB sample of the same model's CRIU images is 24.7% zero 4 KiB pages, all in runs of at least 64 MiB, in one tail.

Problems:

• The page residency scan (Util::scanOccupiedRangeBatch) skips only absent pages. A page that is present but zero (a pinned buffer larger than its contents, a cleared pool) is written to the image and read back on every restart.
• The non-zero rest of these images does not compress (zstd 0.977 of it, LZ4 worse), so compressing images cannot replace skipping zeros, and decompression costs restart CPU.
• If zero runs were skipped, they would cut a large area into runs smaller than the 1 GiB a payload needs for the parallel readers (restart_parallel_read), and those runs would be read by one thread: a regression on exactly the images that benefit.

Acceptance:

• The writer records resident all-zero runs of private anonymous memory as DMTCP_ZERO_PAGE children, with no image format change: existing restarts read new images, new restarts read old ones.
• Ordinary data costs no measurable checkpoint time; resident pages are left in place (no MADV_DONTNEED on memory that may be pinned for DMA); an environment switch restores the old images byte for byte.
• With DMTCP_RESTART_READERS, runs between zero runs are still read in parallel, without slowing restarts of images with many small areas, cold or warm.
• CPU regressions verify every byte, protections, image size, parallel use and cross-version restarts; the full suite passes with one and with 16 readers.
• GPU qualification on the H200 worker (pinned memory re-registered by CUDA after restart) before a warmctl release ships it by default.