cuda: fix abort on mixed f16/bf16 + q8_0 KV cache (default build)

[zPirx]

Overview

The runtime kernel selector allows mixed f16/bf16 and q8_0 KV pairs, but in the default build (GGML_CUDA_FA_ALL_QUANTS=OFF) the four vec instances are not compiled and the dispatch list does not list them. The selector picks the vec kernel, dispatch finds no case, and it hits GGML_ABORT at fattn.cu:469 on the first decode token. Prefill runs fine, so pp-only testing misses it.

Reproduce with -fa 1 -ctk f16 -ctv q8_0 (or reversed): prefill runs, crash on the first decode token. RTX 4090 (sm_89), CUDA 12.8, base 471fb4e. A print before the abort showed D_Q=256 K=f16 V=q8_0.

Fix: add the four instance files to the default CMake else() list and the four cases to the non-ALL_QUANTS #else block. ALL_QUANTS=ON already had them.

Same missing-instance abort described in ggml-org#21526 for the (turbo*, f16) pairs; this PR covers the f16/bf16 + q8_0 pairs in the default build, which were still missing.

Additional information

After the fix, both combos decode instead of aborting (base 471fb4e, commit 53cd9f4):

=== DECODE K=f16 V=q8_0 ===
| model                          |       size |     params | backend    | ngl | type_v |  fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -----: | --: | --------------: | -------------------: |
| qwen35 27B Q4_K - Medium       |  16.67 GiB |    27.32 B | CUDA       |  99 |   q8_0 |   1 |           tg128 |         38.51 ± 0.15 |
| qwen35 27B Q4_K - Medium       |  16.67 GiB |    27.32 B | CUDA       |  99 |   q8_0 |   1 |  tg128 @ d32768 |         33.60 ± 0.01 |
=== DECODE K=q8_0 V=f16 ===
| model                          |       size |     params | backend    | ngl | type_k |  fa |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | -----: | --: | --------------: | -------------------: |
| qwen35 27B Q4_K - Medium       |  16.67 GiB |    27.32 B | CUDA       |  99 |   q8_0 |   1 |           tg128 |         38.74 ± 0.17 |
| qwen35 27B Q4_K - Medium       |  16.67 GiB |    27.32 B | CUDA       |  99 |   q8_0 |   1 |  tg128 @ d32768 |         35.05 ± 0.01 |

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - AI helped locate the missing instances and helped me write this description clearly (not a native speaker). I ran all builds and tests myself, verified the results and take responsibility for the change.