persimmon : simplify rope logic

[Galunid]

Make it similar to stablelm now that gpt-neox is supported for all layers.
I tried to allow for full offloading, but tmpqkv are on CPU and not on GPU.

[Galunid]

This is possible thanks to #4156

[ggerganov]

You can offload tmpqkv by adding an entry in the k_offload_map:

llama.cpp/llama.cpp

Lines 5262 to 5271 in 3cb1c34

	{ "bqkv", OFFLOAD_FUNC_KQ },
	{ "wqkv_clamped", OFFLOAD_FUNC_KQ },
	{ "tmpk", OFFLOAD_FUNC_KQ },
	{ "tmpq", OFFLOAD_FUNC_KQ },
	{ "tmpv", OFFLOAD_FUNC_V },
	{ "Kcur", OFFLOAD_FUNC_KQ },
	{ "Qcur", OFFLOAD_FUNC_KQ },
	{ "Vcur", OFFLOAD_FUNC_V },