fix: include create_exllama_buffers and set_device for exllama

[drbh]

This PR reimports set_device and set_device which are later used in Warmup

text-generation-inference/server/text_generation_server/server.py

Line 104 in 9a7830b

create_exllama_buffers(request.max_prefill_tokens)

and the buffer is used here

text-generation-inference/server/text_generation_server/layers/gptq/exllamav2.py

Line 214 in 9a7830b

output = ext_gemm_half_q_half(x, self.q_handle, self.outfeatures, force_cuda)

This should resolve the ex2 test cases that are receiving a None instead of the pointer to the buffer

[drbh]

merging as this fixes imports that were accidentally removed in #2262 and only adds the missing import lines