Add OpenAI-compatible agent profile

[WarGloom]

This is the focused follow-up from the closed local-agent setup PR. It only adds the generic openai_compatible profile path, so BrainDB can point the internal agent at OpenAI-compatible /v1 endpoints without adding another provider abstraction.

What changed:

• add openai_compatible to _LLM_PROFILES and let AGENT_BASE_URL override the profile base URL
• pass AGENT_BASE_URL and AGENT_API_KEY into the API container
• document the env vars with an Ollama-style example using the generic profile name
• add unit coverage for the new profile and existing profile regressions

What this does not include:

• no local_ollama alias
• no Codex profile or OPENAI_API_KEY passthrough
• no fast-path save routing
• no bundled Postgres, manage script, or GPU compose changes

Validation:

• python -m pytest tests/test_config_profiles.py -q -> 7 passed
• docker compose config
• python -m py_compile braindb/config.py tests/conftest.py tests/test_config_profiles.py
• manual Settings smoke with LLM_PROFILE=openai_compatible, AGENT_MODEL=openai/llama3.2:3b, AGENT_BASE_URL=http://localhost:11434/v1

I could not run the full suite in this shell because the local Python environment is missing agents and psycopg2; the targeted config tests above do not depend on those packages.

[dimknaf]

Thanks — direction is great, two scoping nits before merge:

1. Rename to follow the provider-prefix convention.

Existing env vars are DEEPINFRA_API_KEY, NVIDIA_NIM_API_KEY, VLLM_API_KEY — provider-name prefix. AGENT_* reads as "auth for calling BrainDB's agent endpoint", which these aren't. They're the LLM provider's credentials. Suggest:

• AGENT_API_KEY → OPENAI_API_KEY
• AGENT_BASE_URL → OPENAI_BASE_URL

Bonus: every OpenAI-compatible tool (Ollama, LM Studio, copilot-api, vLLM) already calls them this, so it's instantly recognisable.

2. Scope the base_url env override to the new profile only.

Today no profile lets you override base_url via env — they're all hardcoded in _LLM_PROFILES. The current diff makes AGENT_BASE_URL a global override that would silently re-point vllm_workstation, nim, etc. too. No existing profile needs that capability — only openai_compatible does.

Concretely:

@property
def resolved_base_url(self) -> str | None:
    if self.llm_profile == "openai_compatible":
        return self.openai_base_url or None
    return _LLM_PROFILES[self.llm_profile].get("base_url")

That keeps every existing profile's resolution path untouched, gives the new profile the env-driven flexibility it needs, and the 2-line docker-compose.yml addition (OPENAI_API_KEY + OPENAI_BASE_URL env passthroughs) is fine because it follows the existing explicit-allowlist pattern.

With those two changes I'd merge happily — the rest of the PR is good (regression coverage for existing profiles is a particularly nice touch).