#1 Drop $enable flag on skipDuplicates() scope guard
The $enable param made every non-skipDuplicates createDocuments call pay
for a closure allocation + extra function call. Branch at the call site
instead so the cost only applies when the flag is actually set.

  - Adapter::skipDuplicates(callable, bool) → skipDuplicates(callable)
  - Database::skipDuplicates(callable, bool) → skipDuplicates(callable)
  - Database::createDocuments, Mirror::createDocuments, Pool::delegate,
    Pool::withTransaction now branch inline.

#2 Drop fetchExistingByIds helper, inline find()
The helper's per-tenant grouping defended a hypothetical multi-tenant
batching scenario that no caller exercises (relationships are intra-
tenant, callers always batch per tenant). Existing patterns in the same
file (refetchDocuments, relationship loading) just call find() directly.
Match that idiom and drop ~70 lines.

#4 Mirror: only capture inserted docs in skipDuplicates mode
The captureOnNext accumulator paid the cost (closure + per-doc array
push) on every createDocuments call, including the common non-skip path.
Branch at the entry of Mirror::createDocuments so the capture only
happens when skipDuplicates is set; the non-skip path passes through
to source/destination unchanged.

#5 Move getInsertKeyword/Suffix/PermissionsSuffix to getters cluster
Were sitting next to createDocuments(); moved to the getSupport*
cluster around line 1030 where other adapter-capability shims live.

Not addressed:
- #2 partial: the existing patterns (refetchDocuments etc.) don't handle
  tenant-per-document multi-tenant batches either, so this is consistent.
- #3 (drop the pre-filter): rejected. createDocumentRelationships runs
  in the encoding loop BEFORE the adapter's INSERT IGNORE no-ops the
  parent, so dropping the pre-filter would deterministically duplicate
  child rows on every CSV re-import of a collection with relationships
  (not a race window — every call). The relationships test verifies
  this. Reverting would require reintroducing the deferred-relationships
  scaffolding we just removed, and the adapter still couldn't tell us
  which parents were actually inserted (SQL INSERT IGNORE has no per-row
  reporting). Pre-filter stays.