v0.2.3

fix(llm): make request timeout configurable; document local model sizing

Add AURSCAN_TIMEOUT (seconds, default 180) so slow CPU-only local
backends aren't killed mid-generation by the previously hard-coded
deadline. Surface actionable guidance instead of opaque "context
deadline exceeded", give each fallback URL its own full budget, and
send max_tokens on the OpenAI call.

Document which local models are suitable (size-vs-suitability table,
VRAM guidance) and the two settings users get wrong: num_ctx and
AURSCAN_TIMEOUT.

Closes #8, #1

Signed-off-by: Andreas Reichel <andreas@manticore-projects.com>