add build resources in trt-llm config

[michaelfeil]

🚀 Background

• current engine builder jobs need more GPU memory CPU memory and CPUs in the build job than in the deployment.
• Generally speaking, the number deployment typically needs only half of the gpus that are needed to set during build time. number_builder_gpus: int = 2*deployment_gpus

For some models, we also need more CPU memory for the build stage, especially in cases where CPU memory * 2 < VRAM.

Examples:
For e.g. a 70B model from fp16 to fp8 we need around:

• 140GB Vram
• 140GB (loading the model) (potentally + 70GB cpu memory for export.)
  -> H100:1 has currently 234GB - works

L4:1 and A10G:1 have issues, as they have issues with memory.

This PR is just a draft, there could potentially be a much better way to implement this, e.g. by finding the actual replica and get the request limits from it, instead of "blindly" doubeling.

💻 How

🔬 Testing

🚢 Release requirements

• Pre-release: Specify here
• Post-release: Specify here
• External communication: Specify here