Support GPU & DRF

[k82cn]

Motivation

Currently, the faire share policy only support cpu and memory. When integrating flame with SGLang/vLLM/Pytorch, the GPU resources are requried. This feature is used to support GPU resources in Flame for inference and training workload.

Function Specification

1. The executor manager will report GPU resources accordingly
2. The session manager will support both faire-share by slots and DRF policy
3. The SDK can request GPU resources

Solutions

1. Support GPU in slots parsing, so the faire-share policy will support GPU
2. Introduce a new DRF policy for GPU

Additional context

N/A