ggml : fix for non-virtual destructor issue of gguf_writer_base

[yymin1022]

Overview

gguf_writer_base declares virtual member functions but its destructor
was non-virtual. Deleting a derived writer instance through a
gguf_writer_base* (or any base-class pointer/reference) skips the
derived class's destructor, because it is not virtual.

This can leak resources owned by the derived writer and is undefined behavior per the C++
standard when the dynamic type differs from the static type at the
point of deletion.

Change

Add virtual to gguf_writer_base's destructor in ggml/src/gguf.cpp.

-    ~gguf_writer_base(void) = default;
+    virtual ~gguf_writer_base(void) = default;

Additional information

Found while cross-compiling ggml/llama.cpp as a static library inside
an AOSP tree (vendor partition), which enables
-Werror=non-virtual-dtor unconditionally by AOSP build rule.

Confirmed the class has no other polymorphic-deletion issues after the fix; build completes
cleanly with this flag enabled.

No functional change to existing single-inheritance, base-pointer-free
usage; this only affects correctness when a derived writer is deleted
via a base pointer, which is the intended usage pattern given the
virtual methods already present on the class.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES (Claude Sonnet5, Just debug for my AOSP build issue, and this PR is fuilly written by myself)

[yymin1022]

@JohannesGaessler Hello, could you please take a look when you get a chance?

[JohannesGaessler]

I've given this PR priority since you went so far as to email me about it but this was absolutely not warranted on your end. The gguf_write_base class is only visible in gguf.cpp and it is never used as a pointer where this UB would surface. So this PR, while it avoids a potential bug in the future, is not at all urgent. Don't send me emails like this again or I will ban you.

@ggml-org/maintainers can I please get a second approval?