[FEAT]: Wire durable xdist shard transport into the pytest plugin

[nina-msft]

Description

Stacked on microsoft#113 (dev/nina-msft/xdist-shard-primitives).
This intra-fork draft targets the PR-A branch so the diff shows only the wiring
added on top of the shard primitives. Once microsoft#113 merges into
main, this branch will be rebased onto main and reopened against
microsoft:main.

This PR wires the durable shard-transport primitives added in
microsoft#113 into the pytest plugin, closing the two durability gaps
@bashirpartovi asked us to "call out clearly" in the review of
microsoft#73:

• Gap [FEAT]: Wire durable xdist shard transport into the pytest plugin #3 — a worker killed mid-run lost every already-finished result. The
  inline workeroutput transport only flushed at a clean pytest_sessionfinish.
• Gap [META]: Add GitHub Related Files microsoft/RAMPART#4 — the size cap discarded a worker's whole payload when the aggregate
  exceeded --rampart-xdist-max-bytes, not just the one oversized transcript.

What changed

For local popen topologies (plain pytest -n N, or --tx where every gateway
is popen), each worker now streams every finished result to its own on-disk
JSONL shard, flushing per write. The controller reads the shards back in
pytest_testnodedown. workeroutput carries only a small completion sentinel
(schema, expected record count, trial specs).

• Durable: a worker killed mid-run (crash, OOM, timeout, -x) keeps every
  result it had already flushed; the controller recovers them from the shard and
  marks the run incomplete rather than losing the worker's whole contribution.
• Granular size cap: the cap now applies per record — one oversized result
  is dropped as a truncation marker while its siblings survive.
• Remote fallback: any non-popen gateway (ssh, socket, via proxy)
  transparently falls back to the existing inline transport, since the controller
  cannot read a worker-local shard file. Eligibility is gated by shard_eligible.

Controller/worker wiring: pytest_configure provisions the shared shard dir on an
eligible controller; pytest_configure_node pushes it into each worker's input;
the worker opens a ShardWriter; the _rampart_collect fixture appends per
result; pytest_sessionfinish closes the shard and writes the sentinel;
pytest_testnodedown recovers and merges. Shard dirs are removed at run end
unless --rampart-keep-shards is passed. Single-process and remote runs are
unchanged.

Breaking changes

None. Behavior is identical for single-process runs and for remote/proxied xdist
topologies. Local -n runs gain durability transparently; the only observable
difference is more granular incomplete-reason strings (per-record rather than
whole-payload).

Checklist

• pre-commit run --all-files passes
• Tests added or updated for changes
• Documentation updated

---

Note: Keep this PR in draft until microsoft#113 merges. Once it
lands on main, rebase this branch onto main and reopen against
microsoft:main, then mark ready for review.