ggml: new gpu kernels + extends ggml_leaky_relu + ggml_pad (#621)

ggml: new gpu kernels + extends ggml_leaky_relu + ggml_pad (#621)

* add new cuda kernels and new op ggml_pad

* add ggml_tanh cuda kernel

* remove old broadcast impl

* restore some changes

* cuda: optimized im2col + group_norm kernels

* extent ggml_leaky -> ggml_leaky_relu

* fix some code issues

* cuda: concat support 4 dims

* cuda: fix ggml_acc + add backends ops test

* restore ggml_pad + add backend op test

* metal : implement GGML_OP_ACC

* ggml : fix bug in ggml_upscale

* metal : add ggml_upscale

* metal : add ggml_tanh

* metal : add ggml_gelu_quick

* ggml : make ggml_pad more general purpose

* metal : add ggml_pad

* ggml_leaky_relu as regular op + fix identation

* cuda: ggml_acc admit all op_parms

* negative_slope better pass param

* metal : add ggml_leaky_relu

* metal : add ggml_group_norm

* cuda : minor

* ggml : add GGML_OP_LEAKY_RELU to ggml_compute_backward

* metal : soft max, tanh, supports_op fixes

* test-backend-ops : add sentinels between tensors to detect overflows

---------

Co-authored-by: Georgi Gerganov <redacted>
Co-authored-by: slaren <redacted>