…(R2.5.7)

Implements the only intentional cross-mechanism dependency in the
declarative master-switch design (PR-B): the read-path transient-view
drain mode follows the master switch.

  master ∈ {compression, off}  →  RESTORE mode (existing behavior).
                                   Pointer-swap back to compressed.
                                   Cold values stay compressed across
                                   iterations (R2.5.6 preserved).
                                   ZERO recompression cost.

  master == decompression      →  PERMANENT-DECOMPRESS mode.
                                   Release the saved compressed buffer
                                   (decRef dict frame-ref + reverse
                                   install accounting + zfree); the
                                   temp sds stays installed as
                                   val_ptr; encoding stays RAW.
                                   The robj is now permanently
                                   decompressed.

Without the dispatch, every read-touched key would oscillate
between compressed and uncompressed each iteration in master=
decompression — defeating the operator's drain intent. With it, a
single sweep across the keyspace plus normal read traffic fully
decompresses the dataset. Documented in design §2.5.7.

The drain reads `server.compression_master_switch` once per
beforeSleep invocation; mid-iteration master-switch changes resolve
at the next iteration. The "stale slot" branch is mode-independent:
when the kvstore no longer references our pinned robj, we always
discard (free both buffers + drop pin + free entry).

Refactor: the per-entry RESTORE and PERMANENT-DECOMPRESS finalize
logic is extracted into static inline helpers
(restoreTransientEntry / permanentlyDecompressTransientEntry)
alongside the existing discardTransientEntry. Same shape, easy to
read at the call site, and reusable by the new test-only drain
variants.

Test infrastructure additions:

  testOnlyCompressionDrainTransientViewAsRestore          (new)
  testOnlyCompressionDrainTransientViewAsPermanentDecompress (new)

Both follow the existing AsDiscard pattern: iterate the side-map,
apply the per-entry finalize, skip the kvstore lookup. Lets unit
tests verify the drain side effects without setting up a populated
kvstore.

New gtest cases:

  RestoreDrainPutsCompressedFormBack
    - Materialize, drain in RESTORE mode, verify the robj is back to
      OBJ_ENCODING_COMPRESSED with the original compressed buffer
      pointer reinstated.

  PermanentDecompressDrainReleasesCompressedBuffer
    - Materialize, drain in PERMANENT-DECOMPRESS mode, verify the
      robj is permanently RAW with the temp sds still installed and
      the decompressed bytes round-tripping the original source.

  PermanentDecompressDrainHandlesMultipleEntries
    - Three entries in the side-map, drain processes all of them;
      iteration handles >1 entry without leaks or state confusion.

  RestoreDrainResetsPerIterationBudget
    - Tight savings cap: materialize → drain (restore) → materialize
      again. The second materialize must succeed (would fall back to
      permanent-decompress if the budget hadn't reset).

Verified locally:

  make -j2 -C src SERVER_CFLAGS=-Werror   # clean (BUILD_ZSTD=yes)
  make -C src distclean
  make -j2 -C src SERVER_CFLAGS=-Werror BUILD_ZSTD=no  # clean
  make -j2 -C src test-unit
    367 tests pass (was 363 — 4 new for the two-mode drain).
  ./runtest --single unit/type/compression  # 13/13 pass

What's NOT in this PR (PR-C3 follows): sweeper engine + COMPRESSION
SWEEP FORCE command. The transient-view drain in master=decompression
is now correct on its own (one sweep would drain the keyspace via the
read path), but the explicit sweeper is the more direct convergence
mechanism documented in R2.1.2 and is the natural pair to this PR.