fix(copy): apply --connection-mode to CopyOnMaster path

[talmacschen-arch]

fix #32

---

Change logs

--connection-mode pull was advertised as "destination connects to source",
but the CopyOnMaster strategy ignored the flag and always made the source
master dial the destination master. Two table classes are forced into that
strategy — DISTRIBUTED REPLICATED tables (hardcoded) and tables with
rows <= --on-segment-threshold (default 1,000,000) — so in any topology
where the destination can only be reached as a dialer (e.g. destination
cluster running inside Kubernetes, source outside), cbcopy hits one of
those tables and dies with a TCP failure or, in the special case of
pull + small/replicated tables, with relation "public.cbcopy_ports_temp_onmaster_<ts>" does not exist.

This PR makes --connection-mode apply uniformly to every copy strategy.

What changes

• copy/copy_command.go — add the if ConnectionMode == ConnectionModePull branch to CopyOnMaster.CopyTo / CopyFrom (symmetric to the existing CopyOnSegment branches), and have FormMasterHelperAddress pick --source-host vs --dest-host by mode. The SQL-building bodies are split into unexported formatCopyToCommand / formatCopyFromCommand so they are unit-testable without a DB.
• copy/copy_operation.go — drop the || op.command.IsMasterCopy() clause from the three branch points in CopyOperation.Execute. After the change above, the listener side under pull + CopyOnMaster is src just like every other pull strategy, so the master path no longer needs its own override. The IsMasterCopy() call at line 89 (which decides whether to query the _onmaster_ or _onall_ port temp table) is intentionally retained — it is orthogonal to connection mode.
• copy/copy.go and copy/copy_port_helper.go — unchanged. The existing initializeClusterResources already picks srcManageConn as the listener under pull, which is now correct for both _onmaster_ and _onall_ external tables; the previously-needed "fix the port temp table location" turns out to fall out for free.

Behavior matrix after this PR

connection mode	strategy	listener	dialer	port temp table on	net path required
push	CopyOnMaster	dest master	src master	dest	src → dest:5432 + helper ports
push	CopyOnSegment / ExtDest*	dest segs	src segs	dest	same
pull	CopyOnMaster	src master	dest master	src	dest → src:5432 + helper ports
pull	CopyOnSegment / ExtDest*	src segs	dest segs	src	same

The previously-broken cell is the bolded one.

User-visible doc note

Under --connection-mode=pull, --source-host must be an address reachable from destination cluster nodes (a 127.0.0.1 literal works only when the destination sees itself as the same host). This is the symmetric constraint to push mode's --dest-host, and is now spelled out in the flag's --help text and in the README's "Hostname Resolution" subsection.

Tests

• Unit (copy/copy_command_test.go, new) — table-driven, asserts the SQL emitted by CopyOnMaster in all four (push/pull) × (CopyTo/CopyFrom) quadrants. Each case checks both required tokens (--listen, --host, --port, --data-port-range, --direction send/receive, --cmd-id) and tokens that must be absent on the corresponding branch, so a regression where the two branches build identical strings would fail loudly.
• E2E (end_to_end/basic_test.go, new It block) — runs cbcopy --connection-mode pull with deliberately distinct literal values for --source-host (127.0.0.1) and --dest-host (localhost) that both resolve to the coordinator on the test cluster, forces TEST_COPY_STRATEGY=CopyOnMaster, and grep-asserts the debug log: the sending side must contain --listen and not --host, the receiving side must contain --host 127.0.0.1 (i.e. the source value) and not --listen. The pre-fix code would emit the mirror image and fail at all four token-level assertions. Note: the existing {CopyOnMaster, pull} entry in the table-driven loop did not catch this regression on a single-cluster test bed because --source-host and --dest-host come from the same env-derived value there.
• Manual — verified on a real multi-host SynxDB4 / Cloudberry 2.1.0 cluster (4 hosts) running make end_to_end.

Contributor's checklist

• Clear title and commit messages (4 commits, see git log)
• Signed CLA (prior contribution)
• Read code contribution guide
• Linked to existing issue ([Bug] --connection-mode pull is not applied to the CopyOnMaster path (replicated / small tables fail in dest-only-reachable topologies) #32)

[talmacschen-arch]

Hi @gfphoenix78, sorry to bother you — thanks again for the quick turnaround on #35!

Whenever you have a spare moment, would you mind taking a look at this one as well? It's a small fix scoped to making CopyOnMaster honor --connection-mode pull (previously it always went through push regardless of the flag). I've verified it on a 4-host SynxDB4 cluster end-to-end and added unit + e2e coverage. No rush at all — just flagging it in case it slipped off the radar.