cuda : add support for f16->f16 GGML_OP_SET_ROWS

[fairydreaming]

Overview

This PR adds support for CUDA GGML_OP_SET_ROWS where src0 type is f16 and dst type is also f16. This will come in handy in models based on DeepSeek Sparse Attention (DeepSeek V3.2, V4, GLM 5.0, 5.1, 5.2) where we have to "sparsify" KQ mask based on lightning indexer indices. With f16 support in GGML_OP_SET_ROWS it will be possible to use f16 FA KQ masks without any extra casts, this will reduce memory usage and simplify code.

Additional information

set_rows_cuda() already supported f16 in src_t template argument internally, so I simply added two fully specialized set_rows_cuda<half, int32_t>() and set_rows_cuda<half, int64_t>() that omit unnecessary dst type cases (necessary as set_rows_cuda_quant() handles only float src0 type). Partial specialization would be more compact, but is not supported for functions.

GGML_OP_SET_ROWS tests for f16->f16 case are in #25344 that shall be merged first.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[ggerganov]

I think the HIP quality check is not caused by these changes?