[core] 🚨 Completely remove cache positions (#44181)
* fully remove from the mask api
* fully remove from cache api
* remove it from llama + modulars
* remove them from all modulars of llama descendants
* reapply all those modualrs
* a few more
* style
* style
* better get_seq_length
* fix signatures
* fix device
* win against dynamo
* compatible with cuda graphs
* add args to absorb
* fix lfm2
* fix executorch
* fix zamba2
* fix encoder-decoder
* fix
* fix t5gemma
* oupsi
* oupsi again
* paddleocr
* fix lfms
* fix export
* more export fix
* lfm2
* fix cuda graph super weird issue
* oupsi forgot to add it elsewhere
* fix
* fix all finally
* style
* fix break
* latest model
* style
* add doc
* fix audioflamingo
* doc
* bc on internal mask functions
* oupsi version
* nemotron
* eurobert
