assemble_cvd: avoid repeated sparse super conversions

[zhangxp1998]

Summary

• materialize a sparse build super.img directly as the per-instance raw image before replacing DLKM partitions
• preserve the per-instance raw super.img across resume when the base OS inputs are unchanged
• publish a newly converted image via a temporary file and atomic rename

Why

The previous repack flow copied the sparse build image and invoked lpadd twice. Each invocation converted the full super image to raw and back to sparse, after which assemble_cvd converted it to raw once more. Resume also deleted the resulting per-instance raw image and repeated this work.

Creating the per-instance raw image first lets both lpadd --replace calls operate in place. Keeping that image on resume avoids the initial conversion on subsequent kernel/ramdisk repacks. Existing OS-input invalidation still decides when the preserved image is unsafe to reuse.

Validation

• bazel build //cuttlefish/host/commands/assemble_cvd:assemble_cvd
• local end-to-end A/B using the same 8 GiB sparse super and two DLKM replacement images:
  • previous flow: 10.126 s
  • direct-raw flow: 0.649 s (15.6x faster)
  • final 8 GiB raw images are byte-for-byte identical
  • allocated blocks decreased from 16,777,232 to 3,268,416