flowctl: preview runs on the V2 runtime, remove V1 preview

[jgraettinger]

Replaces the V1-runtime-backed flowctl preview with the runtime-next + shuffle
implementation developed as flowctl raw preview-next, and deletes V1 preview
along with the runtime harness it was the sole consumer of.

raw preview-next is retained as a hidden, deprecated alias onto the same
clap::Args struct, so existing callers — estuary/connectors CI, the in-repo
pytest suite — keep working untouched.

Commits

1. flowctl: preview runs on the V2 runtime — re-point Command::Preview,
   hide + deprecate the raw preview-next alias, delete
   flowctl/src/preview/{mod,journal_reader}.rs (−733).
2. runtime: remove the now-dead V1 preview harness — delete
   runtime/src/harness/ and the private exchange combinator (−1131). Split
   out so the first commit proves the harness has no remaining callers and
   this one reviews as a pure deletion.

Behavior changes

The flag surface is a strict superset (--shards and --debug-port are
additive; nothing is removed). Three behaviors do change:

• Derivation stdout gains ["collection/name",<doc>] framing, which
  captures and materializations already used. Stripping the wrapper reproduces
  the old output exactly. Nothing downstream parses it. The three site/
  derivation tutorials show bare documents and are now stale — but site/ no
  longer serves docs.estuary.dev (90e8261), so that correction belongs in
  estuary/docs and is deliberately not made here.
• --delay with --fixture is now an error. A fixture's transaction
  boundaries are fixed by its own commit markers, so a raised minimum
  transaction duration has nothing to batch.
• Unset --sessions no longer equals --sessions -1 for captures. Unset
  holds an EOF'd session through the capture's poll interval as production does;
  -1 stops at EOF (f259b8f).

Why the module isn't moved to preview/

flowctl preview is backed by a module still named raw/preview_next/. That is
deliberate. #3416 renames five of that module's files out to a new
crates/runtime-local crate and rewrites the rest, so moving now means
rename-vs-rename conflicts on every one — and would land the wrong layout
regardless, since after that split the code deserving the name preview/ is only
the flowctl-specific residue.

git merge-tree against the rebased #3415–#3419 stack reports zero conflicts
in either direction. The only file both touch is raw/mod.rs (this PR edits the
PreviewNext arm; #3417 adds a Test arm beside it), and the merged result was
checked by hand to be semantically correct.

A TODO(johnny) on raw::Command::PreviewNext records the deferred move, the
in-repo callers that must migrate first (tests/preview/*.py and the
tests/soak/** comments, both introduced by #3416), and the eventual removal of
the alias.

Verification

• Both crates build clean under mise with no unused-code warnings — the
  compiler is the proof the harness is dead.
• flowctl preview --help carries the original doc text with the V2 flag set.
• raw --help no longer lists preview-next; raw preview-next --help still
  dispatches and reports itself deprecated.

Note that the tests/*.py snapshot suite could not verify anything: it is not
wired into CI, and has been broken since 85253a1 swapped the fixture reader
without converting the checked-in fixtures. The framing claim above was instead
verified directly against the V1 and V2 snapshot artifacts.

[williamhbaker]

LGTM