feat: Add feature gate

[yeahdongcn]

This PR introduces a feature gate mechanism in ktransformers, providing centralized control over specific features. This helps streamline feature management and reduces the frequent invocation of get_compute_capability.
Vendors can extend this file to implement more fine-grained controls.

Testing Done

$ python ./ktransformers/local_chat.py --force_think true --cpu_infer 64 --model_path deepseek-ai/DeepSeek-R1 --gguf_path /home/gpuserver/models/DeepSeek-R1-Q4_K_M --port 10002
flashinfer not found, use triton for linux
Feature gate initialized: KTRANSFORMERS_USE_TORCH_NATIVE=False, KTRANSFORMERS_USE_FLASHINFER=False
using custom modeling_xxx.py.
using default_optimize_rule for DeepseekV3ForCausalLM
Injecting model as ktransformers.operators.models . KDeepseekV2Model
...

[yeahdongcn]

#787 is no longer necessary.

[yeahdongcn]

Rebased from upstream/main.