ggml-cuda : compute ptrs for cublasGemmBatchedEx in a kernel

[slaren]

Computes the pointers in kernel so that this can be done asynchronously, and uses the CUDA pool to avoid the call to cudaMalloc.

Fixes #3884

[slaren]

Ideally, we would also launch cuBLAS from the same kernel, but it seems that using device-side cuBLAS would require some changes to the build system.

model	size	params	backend	ngl	test	t/s
mistral 7B mostly Q8_0	7.17 GiB	7.24 B	CUDA	99	pp 512	3673.15 ± 55.61
mistral 7B mostly Q8_0	7.17 GiB	7.24 B	CUDA	99	tg 512	79.59 ± 0.34

build: a9ab02e (1456)

[ggerganov]

Tested and it also works on the GTX 1660 which was crashing when using the mem pool + memcpy