Continuous batching: find the sharded key/value projections under any decoder path (#50366)
* Continuous batching: find the sharded key/value projections under any decoder path
A vision-language model's TP plan keys start with model.language_model., so the engine took its key/value heads as
replicated and sized the cache with all of them on every rank: Qwen3.5 at tp=2 fails on the first decode step.
* Nit
---------
Co-authored-by: remi-or <remi.pierre_o@orange.fr>