…#71)

WARNING: This feature is brand new and has not been tested beyond the
existing regression suite.  It should be thoroughly tested — including
stress, power-loss, and wear-levelling scenarios — before being used in
any production system.

When LX_NAND_FLASH_ENABLE_LAZY_SECTOR_RELEASE is defined and free blocks
exceed LX_NAND_FLASH_SECTOR_RELEASE_LAZY_THRESHOLD (default 10), releasing
a sector from a full block defers the copy+erase.  A tombstone page is
written to a freshly-allocated block; the old full block is flagged
COMPACTION_PENDING in its block status and recorded in a new per-LG
compaction table.  Compaction (merge + erase) is triggered lazily:

  * sector_write: before writing to a full block with a pending source
  * block_data_move: intercepts wear-levelling moves for pending LGs
  * defragment: iterates all pending LGs (now actually implemented)
  * block_allocate: emergency compaction when the free list is empty

Reads transparently fall back to the compaction-pending block when the
primary block does not contain the requested sector.

Crash recovery is handled in open_extended:
  - Phase 1: detects COMPACTION_PENDING blocks and either clears the flag
    (crash before mapping update — abort) or rebuilds compaction_table
    (crash after mapping update — resume deferred compaction).
  - Phase 2: erases orphaned ALLOCATED blocks not present in any mapping.

Without the flag, all paths remain unchanged and defragment returns
LX_NOT_SUPPORTED as before.

Fixes: #70

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>