CUDA: fix q_nope_absorbed precision for Deepseek 2 Lite f16

[JohannesGaessler]

I noticed that Deepseek 2 Lite when calculating perplexity on Wikitext was returning worse results for FP16 weights than with q4_0 weights: with 10 512 token chunks FP16 resulted in 27.6215 while q4_0 resulted in 8.1775. The problem seems to be numerical issues in the calculation of q_nope_absorbed, specifically with CUDA and batch sizes > 1. If FP32 precision is used the perplexity becomes 7.9094.

On master ggml_cuda_mul_mat does not always respect the precision set via ggml_mul_mat_set_prec. ggml_cuda_mul_mat_batched_cublas only supports FP16, FP16 -> FP16 GEMM but is used regardless of the requested precision. This PR makes it so that if higher precision is requested ggml_cuda_op_mul_mat_cublas is used instead (which supports FP32 precision). Long-term I think we should aim to remove ggml_cuda_op_mul_mat and refactor the cuBLAS code. I'm currently working towards the former; what I think is specifically needed is MMQ support for batched and non-contiguous inputs and backend-agnostic support for tensor parallelism.

[jukofyork]

Yeah, q_nope_absorbed is basically doing the KQ multiplication so likely it will suffer from the same overflow problems - I had similar problems when I tried to use FP16 for it too.