perf(sync-service): batch cold shape startup work

[joshdchang]

Summary

This is stacked on #4588 and targets its branch so the inspector-cache work can be reviewed separately.

When many shapes for previously unseen relations start together, they currently serialize repeated catalog and publication work. This change:

• loads cold relation metadata in batches rather than issuing one catalog query per relation
• batches publication additions and replica-identity updates in one transaction
• tracks the in-flight relation set so overlapping refreshes do not resubmit the same work
• falls back to the existing per-relation path if a batch update fails

It does not increase startup or snapshot timeouts.

Evidence

On local PostgreSQL 18 with 40 distinct one-row tables requested concurrently from a cold Electric instance:

Runtime	Wall	Median	p95	Max
upstream 1.7.3	0.495s	0.404s	0.479s	0.485s
this branch	0.171s	0.129s	0.168s	0.169s

All 40 requests returned HTTP 200. A repeated warm fanout completed in 0.024s wall time.

Validation on the latest-main integration branch:

• 1,843 tests, 391 doctests, and 9 properties; 0 failures
• focused relation, publication, router lifecycle, PgBouncer, and streaming coverage passes

Validation on this stacked branch:

• 92 focused tests; 0 failures

Review notes

The latest-main integration branch also contains post-#4588 connection-error normalization that this stacked change relies on. Once #4588 is rebased or merged, this PR should be rebased and retargeted to main before merge.