add reasonning budget parameters in chat handlers

[teux91]

could we have the option to limit the number of tokens generated for thinking via parameters equivalent to
--reasoning-budget and --reasoning-budget-message in llama.cpp CLI in the chathandler or elsewhere?

Very useful to control the thinking effort of the models , feel gemma 4 and qwen 3.6 tend to overthink in thinking mode

[JamePeng]

This will likely require building a CustomSampler for ReasoningBudgetSampler to support this feature, which will probably have to wait until I'm done with the exhibition.

[teux91]

Noted, take the time you need