cuda: fix MSVC duplicate-COMDAT link failure from twice-nested retired_buf

[tmark00]

What

feature/turboquant-kv-cache at 80007e715 does not link on Windows/MSVC. q8_cache and
tq_rot_cache are unnamed member structs of ggml_backend_cuda_context, and each nested its own
identical struct retired_buf:

    struct {
        ...
        struct retired_buf { char * ptr; size_t cap; int dev; };
        std::vector<retired_buf> retired;
    } q8_cache;

    struct {
        ...
        struct retired_buf { char * ptr; size_t cap; int dev; };   // same again
        std::vector<retired_buf> retired;
    } tq_rot_cache;

MSVC decorates a type nested in an unnamed struct as <unnamed-tag>, so both nested types mangle
to the same decorated name. The two std::vector<retired_buf> instantiations then emit COMDAT
sections with identical names into a single ggml-cuda.cu.obj, and link.exe rejects the object:

ggml-cuda.cu.obj : fatal error LNK1179: invalid or corrupt file: duplicate COMDAT
'??$?0$$V@?$_Compressed_pair@V?$allocator@Uretired_buf@<unnamed-tag>@ggml_backend_cuda_context@@@std@@
V?$_Vector_val@U?$_Simple_types@Uretired_buf@<unnamed-tag>@ggml_backend_cuda_context@@@std@@@2@$00@std@@
QEAA@U_Zero_then_variadic_args_t@1@@Z'

Because ninja deletes the target before linking, a failed build also leaves the tree without
ggml-cuda.dll, so an incremental rebuild on Windows takes the previously working binary with it.

Fix

The two definitions are identical, so hoist a single retired_buf to class scope. As a member of
a named class it mangles unambiguously, and the duplication goes away as a side effect. Every use
is either auto or braced initialisation (ggml-cuda.cu:751-762, mmvq.cu:1300,
mmvq-tq.cu:927), so no call site changes.

Who introduced it

Neither parent is at fault alone: 767b6278e added the q8 quantize cache and 3df06b110 the TQ
rotation cache, each with its own nested copy. The collision only exists once both are in the same
translation unit.

Verification

Windows 11, MSVC 14.51, CUDA sm_120 (RTX 5090 Laptop), Ninja, Release, BUILD_SHARED_LIBS=ON,
CMAKE_CUDA_ARCHITECTURES=120-real:

• 80007e715 unmodified: FAILED: bin/ggml-cuda.dll, LNK1179 as above
• 80007e715 + this commit: full cmake --build completes, 0 errors, ggml-cuda.dll produced

Non-MSVC toolchains were not affected to begin with — GCC and Clang give the nested types distinct
mangled names — so this should be a no-op there.

🤖 Generated with Claude Code

[TheTom]

Thanks, and sorry: that nesting came from my fix on #343, and the hosted Windows jobs build with clang so they never saw it. Hoisting the one retired_buf to class scope is the right fix. CI was waiting on first-contributor approval, now running; merging when it is green.