vulkan: Implement top-k

[jeffbolznv]

Based on #17365. Ready for review, but Draft until that lands.

[ggerganov]

I ran the exhaustive top-k tests on my Mac and it got stuck at some point:

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 83ba67502..035656943 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -7654,9 +7654,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     // exhaustive top_k tests
-    //for (int i = 1; i < 9999; ++i) {
-    //    test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {i, 2, 1, 3}, rand() % i + 1));
-    //}
+    for (int i = 1; i < 9999; ++i) {
+        test_cases.emplace_back(new test_top_k(GGML_TYPE_F32, {i, 2, 1, 3}, rand() % i + 1));
+    }
 
     for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR, GGML_SCALE_MODE_BICUBIC}) {
         test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));

make -j && ./bin/test-backend-ops -o TOP_K -b Vulkan0

...

  TOP_K(type=f32,ne=[2699,2,1,3],k=482): OK
  TOP_K(type=f32,ne=[2700,2,1,3],k=786): OK
^C

So there could be a bug somewhere.

[jeffbolznv]

Yes, there was a bug in the pipeline selection. I'm also looking into a radix-ish implementation to see how the performance is.

[jeffbolznv]

Rebased, this is ready for review.