turboquant: post-merge integration fixes from test validation

Two fixes surfaced by running the full test suite against the squash-merged
turboquant branch, plus one CMake registration.

1. ggml-cuda/ggml-cuda.cu (GET_ROWS supports_op)
   Removed TQ3_1S/TQ4_1S from the CUDA/HIP GET_ROWS supports_op switch.
   TheTom's branch advertised these as supported but never added the matching
   cases to getrows.cu — a latent bug present on both his branch and master.
   master's test-backend-ops triggers it; the scheduler will now route
   get_rows on TQ types to CPU.

2. ggml-cuda/fattn.cu (HIP head-size gate)
   Master's get_best_fattn_kernel falls through to BEST_FATTN_KERNEL_TILE as
   default. On HIP, fattn-tile.cu only instantiates head sizes 64, 128, 256,
   320, 512 (576/640 exceed local memory limits per #ifndef GGML_USE_HIP).
   Without this gate, supports_op returns true for unsupported sizes and the
   dispatch aborts. Now returns BEST_FATTN_KERNEL_NONE on HIP for head sizes
   the tile kernel cannot compile, letting the scheduler fall back to CPU.

3. tests/CMakeLists.txt (test-turbo-quant registration)
   TheTom added tests/test-turbo-quant.c (CPU round-trip diagnostic for
   turbo3/turbo4 quant→dequant→inverse-WHT) but never wired it into the
   build. Registered as a ctest entry linked against ggml + libm.

Test status with these fixes:
- CPU (build-cpu): 51/51 ctest pass, including new test-turbo-quant.
- HIP (build-hip, gfx1151): 50/50 ctest pass with GGML_CUDA_DISABLE_GRAPHS=1
  and test-backend-ops excluded. test-backend-ops itself runs 13674/13677
  internal cases; the 3 remaining failures (CLAMP f16 → inf, bf16 FA graph
  capture) are pre-existing master-side regressions on RDNA3.5+HIP that
  reproduce on plain master and are unrelated to TurboQuant.

ggml/src/ggml-cuda/fattn.cu

@@ -534,6 +534,17 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             }
         }
     }
+#ifdef GGML_USE_HIP
+    // fattn-tile.cu only instantiates head sizes 64, 128, 256, 320, 512 on HIP
+    // (576/640 exceed local memory limits). Without this gate, supports_op would
+    // claim TILE handles e.g. hsk=192/576 and the dispatch would abort.
+    {
+        const int hsk = (int) Q->ne[0];
+        if (hsk != 64 && hsk != 128 && hsk != 256 && hsk != 320 && hsk != 512) {
+            return BEST_FATTN_KERNEL_NONE;
+        }
+    }
+#endif
     return BEST_FATTN_KERNEL_TILE;
 }
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -5248,8 +5248,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
-                    case GGML_TYPE_TQ4_1S:
-                    case GGML_TYPE_TQ3_1S:
                         return true;
                     default:
                         return false;

tests/CMakeLists.txt

@@ -295,6 +295,13 @@ get_filename_component(TEST_TARGET test-c.c NAME_WE)
 add_executable(${TEST_TARGET} test-c.c)
 target_link_libraries(${TEST_TARGET} PRIVATE llama)
 
+# TurboQuant round-trip diagnostic (CPU paths)
+get_filename_component(TEST_TARGET test-turbo-quant.c NAME_WE)
+add_executable(${TEST_TARGET} test-turbo-quant.c)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml m)
+add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}>)
+set_property(TEST ${TEST_TARGET} PROPERTY LABELS main)
+
 llama_build_and_test(test-alloc.cpp)
 target_include_directories(test-alloc PRIVATE ${PROJECT_SOURCE_DIR}/ggml/src)