feat: add ~agent@1.0, ~inference@1.0 and ~sev_gpu@1.0 devices

[jax-cn]

Summary

This PR adds three new HyperBEAM devices along with supporting infrastructure.

---

~inference@1.0 (src/dev_inference.erl)

An OpenAI-compatible inference gateway that routes completion requests to a local or remote LLM backend via relay@1.0.

Key opts read at runtime:

Opt	Default	Description
agent-api-peer	http://localhost:8080	Backend URL
agent-api-path	caller-supplied	API path (e.g. /v1/chat/completions)
agent-api-key	—	Optional Bearer token

Works unchanged with local llama.cpp, OpenRouter, OpenAI, or any OpenAI-compatible endpoint.

---

~agent@1.0 (src/dev_agent.erl)

A ReAct-loop agent device. Iteratively:

1. Calls LLM through inference@1.0/completions
2. Parses tool_calls from the response
3. Dispatches to built-in tools
4. Feeds results back as tool role messages
5. Repeats until the model emits a final answer

Built-in tools: http_request, lookup_data, search_messages, get_arweave_tx, bundle_item

All tools, the LLM call, and system prompt are overridable via Run key parameters.

Architecture

agent@1.0
    └─► inference@1.0 ──► relay@1.0 ──► local llama.cpp
                                    └──► remote provider (OpenRouter / OpenAI)

---

~sev_gpu@1.0 (src/dev_sev_gpu.erl + native/dev_sev_gpu_nif/)

SEV GPU device with NIF support for GPU-backed confidential compute.

---

Supporting changes

• src/hb_opts.erl — register all three devices; add /v1/.* → localhost:8080 route; inference_opts defaults
• src/hb_http.erl / src/hb_http_server.erl — HTTP layer improvements
• src/dev_meta.erl, src/dev_codec_http_auth.erl, src/hb_app.erl — supporting updates
• rebar.config / Makefile — build support for SEV GPU NIF
• docs/devices/agent-at-1-0.md — full device reference
• docs/dev_agent-testing.md — testing / debugging guide
• scripts/agent-test.lua, scripts/e2e-test.sh — test scripts
• mkdocs.yml — add ~agent@1.0 to Devices nav
• .gitignore — ignore .agents/, dev_sev_gpu logs