hexagon: the ggml_hexagon_supported_mul_mat function has incorrect logic when checking quantization types

[zhouwg-jeffzhou]

Overview

the ggml_hexagon_supported_mul_mat function has incorrect logic when checking quantization types

Additional information

fix the below issue:

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - Paired with Trae(Seed-Code). I have reviewed all modified lines and understand the details, tested locally on a Snapdragon 8Elite(aka 8Gen4) phone.

[zhouwg-jeffzhou]

Whether this PR gets approval depends on the maintainer's view:

• It can fix that issue, since the logic inside test‑backend‑ops.cpp differs from normal inference.
• It has no side effect for normal inference.
• It introduces a side effect: all ggml-hexagon's supported quantized mulmat will be filtered out in test‑backend‑ops.cpp. The side‑effect within test‑backend‑ops.cpp is acceptable for ggml‑hexagon, and it is better than having ggml‑hexagon crash when running test‑backend‑ops.cpp.

Why doesn't ggml‑vulkan have this weird issue? The root-cause is that there are certain special requirements in Snapdragon's NPU. I will close this PR accordingly if it can't get accepted within one week.