feat: Add Forge as LLM provider

[Yiiii0]

Description

Adds Forge as an OpenAI-compatible gateway provider in nanobot's provider registry, following the same pattern as AiHubMix.

Changes Made

• nanobot/nanobot/providers/registry.py: Added ProviderSpec for Forge with is_gateway=True, litellm_prefix="openai", strip_model_prefix=True
• nanobot/nanobot/config/schema.py: Added forge field to ProvidersConfig
• nanobot/nanobot/providers/litellm_provider.py: Extended env_extras handling to use os.environ[...] (override) for gateways instead of setdefault, matching the primary env_key pattern at line 60
• nanobot/README.md: Added Forge to provider table

Usage

# In nanobot config:
providers:
  forge:
    api_key: "your-forge-api-key"
    api_base: "https://api.forge.tensorblock.co/v1"  # optional, this is the default

Checklist

• Changes tested locally
• Documentation updated
• Ran pre-commit hooks

I work at TensorBlock and will help maintain this integration.

---

About Forge

Forge is an open-source middleware service for unified AI model provider management. It routes requests across 40+ AI providers with access to thousands of models through a single OpenAI-compatible API.

• Repo: https://github.com/TensorBlock/forge
• Docs: https://www.tensorblock.co/api-docs/overview

[Zongwei9888]

Ported to core/providers in b737060 (via #159), with a Co-authored-by: trailer for you.

Your branch targeted nanobot/, which the v2.0 refactor removed, so the spec was rewritten against core/providers/registry.py — same fields, openai_compat backend instead of the litellm prefix. api.forge.tensorblock.co/v1/models verified reachable (401).

strip_model_prefix=True came across as you had it. That is the field that makes this more than a convenience entry: through the generic custom provider, a habitual anthropic/claude-3 would be sent to Forge verbatim and fail.

Your second change found a real bug here too. The setdefault fix you made in nanobot's litellm_provider.py applies to core/providers/openai_compat.py, which had it only half-right:

if spec.is_gateway:
    os.environ[spec.env_key] = api_key      # gateway wins
else:
    os.environ.setdefault(spec.env_key, api_key)
...
for env_name, env_val in spec.env_extras:
    os.environ.setdefault(env_name, resolved)   # gateway does NOT win

A stale ambient variable outranked the gateway the user had selected. #159 applies the same rule to env_extras, with a test that fails if it is reverted. Latent on main today — only zhipu uses env_extras and it is not a gateway — but it would have bitten the first gateway that needed one. Thanks for catching it.