## What type of PR is this?

- [x] BUG
- [x] Test and CI

## What this PR does / why we need it

This PR fixes two independent nondeterministic test setups without
relaxing production behavior.

### TAE checkpoint replay

TestAppendAndGC2 created its database and tables through low-level TAE
catalog APIs. Those APIs mutate the in-memory catalog but do not append
rows to mo_database, mo_tables, and mo_columns. After a forced
checkpoint truncates the DDL WAL, checkpoint replay can therefore lose
the still-live database.

The prepared-DML replay fence added by #26027 exposed this invalid setup
as ExpectedEOB. The previous version of this PR skipped missing catalog
entries, but diagnostics proved the missing IDs were the live test
database and table. That production-code relaxation was removed.

The test now uses the full DDL helpers and, after restart, opens both
relations and scans every column to verify that each still contains all
99 rows.

### Shard read compatibility test

TestReadValidatesAllRemoteShardsBeforeSending could create its table
before all three CNs reported to the shard server. In the failed Ubuntu
job, all shards were initially allocated to cn1 and balancing the second
shard took 60 seconds. The shared setup context expires after 10
seconds, so the first remote compatibility check returned context
deadline exceeded after one Adjust call instead of reaching the intended
incompatible CN.

The test now waits for every CN before creating shards, bounds every
replica wait, uses separate setup and read contexts, and asserts the
exact incompatibility error. This keeps the test fast and prevents
unrelated timeout errors from satisfying require.Error.

No production code is changed.

## Validation

- TAE focused replay test: count=20
- TAE focused replay test: race count=10
- pkg/vm/engine/tae/db/test full package
- pkg/vm/engine/tae/txn/txnimpl full package
- shard compatibility test: count=30
- shard compatibility test: race count=10
- pkg/shardservice full package
- go build ./pkg/shardservice
- go vet ./pkg/shardservice
- git diff --check

This is the next prerequisite after #26027. It does not expose `ALTER TABLE ... AUTO_INCREMENT=N` yet.

The invariant in this PR is that an AUTO_INCREMENT offset reset and its allocator-cache transition belong to the same transaction:

- the resetting transaction sees its pending offset without publishing it to other transactions;
- statement or transaction rollback retires the private reset cache;
- commit publishes a cache for the allocation epoch actually used by the write;
- stale builders and old cache generations cannot become active after reset/reload/close;
- accepted allocator actions drain during close, while a canceled queued reset cannot mutate the store.

The shared service interface now carries `autoIncrEpoch` and `TxnOperator` through both production allocation callers. `memStore` and `sqlStore` gain monotonic `SetOffset` plus DDL-constrained `ForceSetOffset`; all implementations and mocks are updated atomically.

Out of scope for this PR:

- planner/compiler support for the user-facing ALTER syntax;
- COPY/CLONE allocator reconciliation;
- rolling-upgrade enablement for mixed-version TNs;
- end-to-end BVT for the final SQL behavior.

Diff size, separated for reviewability:

- production: `+1303 / -86`
- tests: `+1832 / -8`
- mocks: `+292 / -18`

Validation:

- `go test -race ./pkg/incrservice -count=1`
- `go test ./pkg/sql/colexec/preinsert ./pkg/sql/colexec/table_clone ./pkg/frontend/test ./pkg/vm/engine/test/testutil -count=1`
- `go test ./pkg/... -run '^$'`
- `git diff --check`
- exact-head `mo-pr-preflight-review`: PASS at `891781d3a927b80b54d38e381ebee0d749c1c589`

Approved by: @XuPeng-SH, @gouhongshen