fix: allocate tmp based on sgmv kernel if available

[drbh]

This PR improves the get_tmp_tensors function to correctly allocate temp tensors based on the sgmv kernel is availability.

This fixes an issue where temp tensors were allocated via the non kernel path, but later checked by a kernel assert.

started

text-generation-launcher \
--model-id meta-llama/Meta-Llama-3-8B-Instruct \
--lora-adapters DavidLanz/Llama3_tw_8B_btc_qlora

req/rep without adapter

curl 127.0.0.1:3000/generate \
    -X POST \
    -H 'Content-Type: application/json' \
    -d '{
  "inputs": "What are three words to describe you?",
  "parameters": {
    "max_new_tokens": 20
  }
}'
# {"generated_text":" (e.g. funny, outgoing, creative)\nI would say that three words to describe me are"}

req/rep with adapter

curl 127.0.0.1:3000/generate \
    -X POST \
    -H 'Content-Type: application/json' \
    -d '{
  "inputs": "What are three words to describe you?",
  "parameters": {
    "max_new_tokens": 20,
    "adapter_id": "DavidLanz/Llama3_tw_8B_btc_qlora"
  }
}'
# {"generated_text":" A. Adventurous, B. Creative, C. Curious\nWhat are three words to describe"}%

[Narsil]

LGTM