Fix mixed-prefix drain in reuse_storage_immediate

[lightsighter]

The oldest-entry drain loop applied the offsets-flavor PendingRelease::release() to every entry it drained, but the allocated/offsets outputs are only meaningful for old_inst's own redistrict entry. A ready follower swept up by the drain misbehaved:

• a plain destroy (empty redistrict_tags) trips the assert(!redistrict_tags.empty()) in debug builds; in release builds split_range returns 0 and clobbers 'allocated', mis-notifying old_inst's already-placed children as ALLOC_EVENTUAL_FAILURE while their tags stay live in current_allocator (leaked ranges)
• a redistrict follower with more children than old_inst overflows the offsets vector inside split_range (out-of-bounds write)

Apply the offsets-flavor release only to old_inst's own entry and drain followers with the void flavor, matching the equivalent drain in release_storage_immediate.

Adds two DeferredAllocBadPathTest units covering both follower shapes; on the unfixed code they reproduce the two assertion failures exactly.

Adjudicated as BUG-7 in tla/allocation/bugs/BUG-7.md (checked in separately with the deferred-allocation verification campaign).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 30.54%. Comparing base (15b3e9b) to head (be8d624).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #475      +/-   ##
==========================================
+ Coverage   30.41%   30.54%   +0.13%     
==========================================
  Files         199      199              
  Lines       41282    41322      +40     
  Branches    14779    14776       -3     
==========================================
+ Hits        12554    12621      +67     
+ Misses      27488    27373     -115     
- Partials     1240     1328      +88     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.