PR #2169: Fix use-after-free in raw_hash_set move-assign with unequal allocators

Imported from GitHub PR https://github.com/abseil/abseil-cpp/pull/2169

## Issue

Fixes #2168 — `raw_hash_set` move assignment into a non-empty table with unequal, non-propagating allocators probes and inserts into the freed backing array (heap-use-after-free, reported by Fengxiaoxx with an ASan trace).

## Root cause

`move_assign(that, std::false_type)` calls `destructor_impl()` on the destination, which deallocates the backing array but leaves `common()` describing the old capacity and control bytes. `move_elements_allocs_unequal()` then calls `reserve()`/`insert()` against that stale state: the insert probe reads freed control bytes (`GroupSse2Impl`/`GroupAArch64Impl` construction → ASan `heap-use-after-free`), writes slots into freed memory, hits the `full table!` assert in debug builds, and hangs in `NDEBUG` builds. With a non-freeing resource (e.g. `monotonic_buffer_resource`) the destination silently keeps its old elements (`dst.size()` = 32 instead of 16).

## Fix

Reset `common()` to the default state immediately after `destructor_impl()`, so the destination is a clean empty table before re-insertion:

```cpp
destructor_impl();
// destructor_impl() leaves common() pointing at the freed backing array
// (or at the stale small-table state), so reset it to the default state
// before re-inserting elements.
common() = CommonFields::CreateDefault<SooEnabled()>();
```

This mirrors the identical reset that `assign_impl()` and `move_elements_allocs_unequal()` already perform on the source's `common()` after teardown.

## Test

New `NoPropagateOnMove.MoveAssignmentWithDifferentAllocToNonEmptyTable` in `raw_hash_set_allocator_test.cc`: move-assigns a 1-element table into a 16-element table under a different allocator id. On the unfixed code the test reports `u.size()==16` and aborts on the `full table!` assertion (`raw_hash_set.h:3498`); the existing `MoveAssignmentWithDifferentAlloc` test only covered an empty destination.

## Verification

- Reporter's `flat_hash_set<int, ..., std::pmr::polymorphic_allocator<int>>` reproducer (two unequal `memory_resource`s, 16 elements each side):
  - Unpatched: ASan `heap-use-after-free` READ in `GroupAArch64Impl` (`hashtable_control_bytes.h:369`) via `find_or_prepare_insert_large` → `insert` → `move_elements_allocs_unequal`; freed region = destination backing array deallocated by `destructor_impl()`.
  - Patched: prints `dst.size()=16`, no sanitizer findings.
- `absl_raw_hash_set_allocator_test`: 29/29 pass (including the new test; it fails and asserts on the unfixed build).
- `absl_raw_hash_set_test`: 297/297 pass.

Merge e47fa80392937a8273d0cc5ced7f5f0260d99653 into 2c256e73951c4209da19359b4347f5035f64add3

Merging this change closes #2169

PiperOrigin-RevId: 986724238
Change-Id: I6103cd3a95d237bc16d771216f05711ea2c32cc8