ROCm and sliding windows fixes

[fxmarty]

Fix models requiring window attention - no need to raise an error at load time in case max_input_tokens < sliding_window.

Also, updates vllm fork commit to use the fixes from ROCm/vllm#28 & fix a few rocm issues

[Narsil]

Fix models requiring window attention - no need to raise an error at load time. Only if context > window length at runtime.

context --max-total-tokens, right ? (We need to crash at load time, crashing randomly at runtime is a terrible UX).

[fxmarty]

@Narsil I assume paged attention always works with sliding window

[Narsil]

LGTM