ESQL: Parallelize newline split discovery for large text files

[julian-elastic]

Problem

FROM <ndjson dataset> | LIMIT 10 over a 9.8 GB S3 file spent nearly all of its
time in planning: split discovery walked the file with ~146 serial blocking
ranged GETs to find newline boundaries, one thread parked on Socket.read. The
cost scales with file size rather than with LIMIT, so no read-side change can
avoid it.

Change

Discovery is now two phases. Planning does no I/O: it either produces splits
directly or emits a descriptor of the probe offsets a file needs. Every probe
across every file is then flattened into one BoundedParallelGather, so a
single global budget bounds in-flight GETs regardless of file count.

• Probe offsets are fixed multiples of the stride, so serial and parallel
  produce identical boundaries; the gather preserves file order. An offset
  that finds no boundary costs only its own split: the splits either side of
  it merge, and every other offset's outcome stands.
• Each probe reads a bounded 256 KB window and drains before closing rather than
  aborting, so the connection returns to the pool for the next probe.
• The effective stride is floored to the 256 KB probe window: below it, probe
  windows overlap the same records anyway, and planning would materialize
  fileLength / stride probe tasks — unbounded below. The floor caps a file's
  probes at fileLength / window.
• Concurrency is capped at 16 and clamped to the node's blob-store permits.
  Without an executor the probe phase walks the same offsets serially on the
  calling thread and produces identical boundaries.
• Cancellation is checked between phases, inside each probe, and after the
  gather joins; a cancelled probe aborts its stream instead of draining it.

Measurements

split_discovery_nanos from the profile, same query and file, 147 splits at the
default 64 MB target split size: 61.0s on main (n=1) versus 2.5s here (n=3,
2.48-3.58s). End-to-end took is not a clean comparison: the read side now
dominates and varies several-fold run to run, which is a separate defect.

Coordinator-local planning change, so no wire format, settings, or cross-node
contract changes. Unit tests cover the probe arithmetic, boundary reduction,
drain-versus-abort, the concurrency ceiling, the stride floor, and cancellation
in each phase.

Closes https://github.com/elastic/esql-planning/issues/1528

[elasticsearchmachine]

Pinging @elastic/es-analytical-engine (Team:Analytics)