Add an ability to not start the agent on some instances

We are currently integrating the Pyroscope Java agent into our infrastructure, and it works great. However, we have encountered a problem: In cases of a large fleet of pods (1000+) and aggressive sampling, we generate too much profiling data for the server to handle. Additionally, the 2-3% profiling overhead becomes a significant resource drain.

I propose adding an option (both in the SDK and as an environment variable) named **PYROSCOPE_CLUSTER_RATIO** with a value in percentage (0-100). This feature would enable the agent only if a randomly generated value in an inclusive range 0-100 **is less than the provided percentage**.

This feature will allow users to save resources on both the server and client sides while still maintaining statistically viable continuous profiling data for large deployments.

I would be glad to help implement this feature if you consider it appropriate.