Add opt-in NDJSON streaming for bulk prediction

[pavsoss]

Summary

Part 2 of 2 for #1021. Adds an opt-in streaming mode to bulk prediction so
large uploads can be consumed incrementally instead of waiting for the entire
result set to be buffered in memory.

Depends on Part 1 (feat/1021-bulk-predict-robust). This branch is stacked
on top of it; once Part 1 merges into the default branch, this diff shrinks
to just the streaming additions.

Behaviour

• Clients opt in with ?stream=ndjson or Accept: application/x-ndjson and
  receive newline-delimited JSON: an opening meta line (model version and row
  count), one result line per scored row, any skipped lines, and a closing
  summary.
• The same per-row isolation and skip reasons apply, so the stream never aborts
  on a single bad row.
• File-type, size, and total-row-cap checks run before the stream opens, so
  those failures still return the normal typed error rather than a half-written
  stream.
• The default buffered JSON response is the fallback and is unchanged; a request
  without the opt-in behaves exactly as before.

Parts

• Part 1 of 2 ([Feature]: Resilient bulk prediction — per-row error isolation, input caps, and reload-consistent version stamping #1021) — resilient, reload-consistent scoring (dependency).
• Part 2 of 2 (this PR) — NDJSON streaming.

Test plan

Coverage for the streaming shape (meta/result/skipped/summary framing and the
model-version header) via both the query param and the Accept header, that the
row cap is enforced before streaming begins, and that the non-streaming JSON
path remains the default.

Closes #1021

[vercel]

@pavsoss is attempting to deploy a commit to the Aditya Sharma's projects Team on Vercel.

A member of the Team first needs to authorize it.