b1703

cuda : fix vmm pool with multi GPU (#4620)

* cuda : fix vmm pool with multi GPU

* hip

* use recommended granularity instead of minimum

* better error checking

* fix mixtral

* use cudaMemcpy3DPeerAsync

* use cuda_pool_alloc in ggml_cuda_op_mul_mat

* consolidate error checking in ggml_cuda_set_device

* remove unnecessary inlines

ggml-ci

* style fixes

* only use vmm for the main device

* fix scratch buffer size, re-enable vmm pool for all devices

* remove unnecessary check id != g_main_device