feat(drivers): support custom HTTP headers for OpenAI-compatible providers

[lifaon74]

Problem

The OpenCode Go gateway (https://opencode.ai/zen/go/v1) now requires an x-opencode-session header on every request (see the official requirements and this write-up on the breaking change). Without it, every completion fails with:

API error (400): {"type":"error","error":{"type":"MissingSessionID","message":"Request is missing x-opencode-session and cannot be routed efficiently."}}

OpenFang's OpenAIDriver already supports extra headers via with_extra_headers() (used hard-coded for GitHub Copilot), but there is no configuration path that populates it, so OpenCode Go subscriptions cannot be used as a provider at all.

Solution

Add an optional headers table to [default_model] and [[fallback_providers]]:

[default_model]
provider = "openai"
model = "glm-5.3-flash"
api_key_env = "OPENCODE_API_KEY"
base_url = "https://opencode.ai/zen/go/v1"

[default_model.headers]
x-opencode-session = "openfang-my-instance"

• Plumbed through DriverConfig and applied in create_driver() to OpenAI-compatible drivers via the existing with_extra_headers().
• Resolved like subprocess_timeout_secs: from [default_model] for the primary provider, from the matching [[fallback_providers]] entry for fallbacks, None otherwise (headers belong to a specific endpoint, not inheritable).
• Values are redacted in Debug output (length only), consistent with api_key handling.
• Docs (docs/configuration.md) and openfang.toml.example updated.

Testing

• cargo check on openfang-types, openfang-runtime, openfang-kernel, openfang-api, openfang-cli, openfang-migrate
• cargo test: 388 (types) + 995 (runtime) pass, including 4 new config round-trip tests
• cargo fmt --check clean, clippy 0 warnings on touched files
• End-to-end verified against the real OpenCode Go endpoint with x-opencode-session set