Context length auto fit (#345)

* Add MLX batched VLM context fitter

* Document MLX context fit calculation

* Clarify context memory budget calculation

* Leave unsupported VLM contexts unchanged

* Best-effort fit compatible VLM caches

* Report fitted VLM context after load

* Fit batched VLM context automatically

* Leave context unchanged when fitting fails

* Skip context fitting for unchunked prefill

* Account for per-layer prompt inputs