[MSan] Destructor poisoning of trivial base includes reused tail padding

LLM Text. The problem is real, although I didn't check it's proposed solution
-----
MemorySanitizer's use-after-destructor instrumentation produces a false
positive when another base subobject occupies the reusable tail padding of
a trivially destructible polymorphic base.

Reproducer: https://godbolt.org/z/v59Yfs4Ea

Compile with:

clang++ -std=c++17 -O0 -g \
-fsanitize=memory \
-fsanitize-memory-use-after-dtor \
-fPIE -pie repro.cpp

Clang 14 exits successfully. Clang 15 and current trunk report
use-of-uninitialized-value inside ReadDuringDestruction::~ReadDuringDestruction().

The object layout is:

- Polymorphic: sizeof=16, data size=12
- ReadDuringDestruction base: offset 12

SanitizeDtorTrivialBase passes BaseLayout.getSize() (16) to the destructor
poisoning callback before ReadDuringDestruction's destructor runs. This
poisons bytes 12–15, even though they contain the still-live base's `value`
member.

Current code:
https://github.com/llvm/llvm-project/blob/main/clang/lib/CodeGen/CGClass.cpp

The instrumentation appears to need the base's data size rather than its
standalone size, or otherwise avoid poisoning reusable tail padding.

Start in clang/lib/CodeGen/CGClass.cpp at SanitizeDtorTrivialBase and reproduce the report with the linked Godbolt example using the provided Clang and MemorySanitizer flags. Trace the size passed to the destructor-poisoning callback and verify that the still-live base member is not poisoned. Done means the reproducer no longer reports a false positive while destructor poisoning remains covered by an appropriate regression test.

Written by the indexing model from the issue text.