Externalize prompts and add llama startup warmup

[vbergeron]

Prompt refactor

• Move all four step prompts (gate, extract, goals, respond) and their few-shot examples out of Prompt module source into priv/prompts/:
  priv/prompts/.txt — system instruction prose
  priv/prompts/_shots.chatml — raw ChatML few-shot turns
• Load both files at compile time via @external_resource + File.read!/1
  so mix compile picks up edits without touching .ex source.
• Expand shot coverage across all four steps:
  • gate: 10 shots covering rhetorical assertions, back-references,
    imperatives-as-queries, social/chitchat disambiguation
  • extract: 11 patterns including graph/edge notation and transitivity
  • goals: 7 patterns covering polar/wh/attribute, superlatives,
    multi-condition decomposition, set-difference
  • respond: 17 evidence shapes (learned fact/rule/constraint/enum,
    true/false/bound queries, unanswered, contradiction, mixed, empty)

Llama server

• Add --mlock to the llama-server argv to pin model pages in RAM after the first warm-up inference pages them in (requires ulimit -l).
• Send :warmup to self once the server transitions to :ready; a Task fires a 1-token completion so the page-fault storm is paid at startup rather than on the first user turn. Failure is non-fatal (logged as warning).