Harden validator exposure and restrict llm endpoint access

[RaoAsim]

Summary

This PR reduces unnecessary validator-side attack surface and adds code-level access controls for the local llm_endpoint.

What changed

• Added PERTURB_ENABLE_VALIDATOR_AXON to make the validator axon optional.
• When validator axon is enabled, it now rejects inbound subnet requests by default instead of exposing an unguarded listener.
• Added IP-based allowlisting to tools/llm_endpoint_service.py.
• llm_endpoint now allows loopback by default and can be expanded explicitly with:
  • LLM_ENDPOINT_ALLOWED_CIDRS
  • LLM_ENDPOINT_ALLOW_PRIVATE_NETWORKS
• Updated env examples and README with the new settings and recommended defaults.

Why

In the current subnet flow, the validator actively queries miners, but there is no obvious subnet-specific need for the validator to accept arbitrary inbound protocol traffic. Keeping the validator axon open without an explicit deny policy increases exposure unnecessarily.

The llm_endpoint was also relying mostly on bind-host discipline for protection. This PR adds application-level client checks so it remains protected even if it is bound to a broader interface by mistake.

Defaults after this change

• PERTURB_ENABLE_VALIDATOR_AXON=false is recommended in validator.env.example
• LLM_ENDPOINT_ALLOWED_CIDRS=127.0.0.1/32,::1/128
• LLM_ENDPOINT_ALLOW_PRIVATE_NETWORKS=false

Notes

• This does not remove the validator axon entirely. Operators can still enable it if their deployment requires it.
• If enabled, inbound validator axon requests are rejected by default.
• llm_endpoint should still be kept on loopback or a trusted private subnet in deployment.

Files changed

• neurons/validator.py
• tools/llm_endpoint_service.py
• scripts/validator.env.example
• scripts/llm_endpoint.env.example
• README.md

Testing

• Verified Python syntax for:
  • neurons/validator.py
  • tools/llm_endpoint_service.py