hexagon: HMX matmul with F16 activation and F16/F32 weights of any row count

[njsyw1997]

Overview

This PR lets the HMX matmul run on F16/F32 weights whose row count is not a
multiple of 32, and on F16 activations. Both show up on tensors that only
exist at runtime, so they cannot be padded or repacked ahead of time the way
quantized weights are; they have to be handled inside the kernels. The typical
case is ggml_conv_2d in the Qwen3-VL vision encoder: it passes the im2col
result as src0 (its row count is the number of patches) and the kernel weight
as src1 (F16 for an F16 mmproj).

It also fixed a typo which will cause the second weight chunk in the batched matmul prologue wrong.

Additional information

git clang-format produced unexpected results. The code is just manually formatted.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes. With help of Claude Farble. All the code has already been reviewed by me.

[njsyw1997]

Close this PR. It's already included in #29974