feat: pre-chunked JSON ingestion with per-chunk metadata and summaries

[ajainvivek]

Summary

• Adds a new POST /v1/tables/{table_name}/ingest/chunks endpoint that accepts an array of pre-split text chunks, each with optional metadata and an optional pre-computed summary
• Chunk ingestion bypasses the extraction and chunking pipeline stages — content goes straight into the tree builder and LLM summarizer
• Per-chunk metadata (page_number, start_line, end_line, section_type, plus any custom keys) is threaded all the way through to NodeMetadata on each PageNode
• BatchSummarizer now skips nodes that already carry a non-empty summary (useful when callers provide pre-computed summaries)
• generate_summaries is always true for chunk ingestion (removed as a caller-facing option) — pre-chunked content almost never includes summaries so LLM generation is the right default
• Docs updated: new "Ingest Pre-chunked Content" section in docs/guides/ingestion.mdx, README diagram and pipeline description updated

Changes

New API

POST /v1/tables/{table_name}/ingest/chunks

Request body:

{
  "title": "My Doc",
  "chunks": [
    {
      "text": "The mitochondria is the powerhouse of the cell...",
      "heading": "Biology Basics",
      "metadata": { "page_number": 1, "start_line": 1, "end_line": 12 }
    },
    {
      "text": "Photosynthesis converts sunlight into glucose...",
      "heading": "Plant Biology",
      "summary": "Plants make food from sunlight via photosynthesis.",
      "metadata": { "page_number": 2, "start_line": 13, "end_line": 28 }
    }
  ],
  "tags": ["biology"],
  "metadata": { "author": "Alice" }
}

Crates modified

Crate	Change
reasondb-ingest	TextChunk, ChunkInput, BuildNode all gain summary: Option<String>; BatchSummarizer skips nodes with existing summaries
reasondb-server	ChunkBodyItem gains summary field; generate_summaries removed from IngestChunksBody/IngestChunksRequest; job processor hardcodes true for chunks

Test plan

• Build passes (docker compose build)
• Container starts healthy
• Chunk without pre-supplied summary → LLM generates summary automatically
• Chunk with pre-supplied summary → summary used as-is, LLM skipped for that node
• Root/parent nodes always get LLM-generated rollup summaries
• page_number, start_line, end_line stored correctly on NodeMetadata