Support setting per-table ANALYZE options in CREATE TABLE and ALTER TABLE

## Feature Request

**Is your feature request related to a problem? Please describe:**

Today the only way to persist an analyze option (buckets, topn, sample num, sample rate,
column choice/list) is to run `ANALYZE TABLE t WITH ...`, which also collects statistics.
That has several consequences:

- Changing one option on a large table costs a full statistics collection nobody asked for.
- Options cannot be declared up front in a schema migration, before the table has data.
- Options are invisible in `SHOW CREATE TABLE`, so schema diffing and code review cannot see them.
- Options live in `mysql.analyze_options` keyed by physical table id, so they are silently lost on
`TRUNCATE TABLE`, on `EXCHANGE PARTITION`, and on BR restore. BR lists `analyze_options` in
`unRecoverableTable` (br/pkg/restore/snap_client/systable_restore.go) precisely because the rows
are keyed by table id.

pingcap/tidb#69853 added `ANALYZE ... WITH DEFAULT ` to clear a persisted option, which
makes the missing other half more visible: you can now reset an option without choosing a value,
but you still cannot set one without collecting statistics.

**Describe the feature you'd like:**

CREATE TABLE t (a int) STATS_BUCKETS = 64, STATS_TOPN = 20;
ALTER TABLE t STATS_SAMPLE_RATE = 0.5;
ALTER TABLE t STATS_BUCKETS = DEFAULT; -- clear, same reset spelling as #69853

No statistics are collected by these statements. The next ANALYZE, manual or auto, picks the
options up.

**Existing half-built scaffolding in the tree (important):**

This is as much "finish or delete dead code" as it is a new feature. Already present and unused:

- Grammar `ALTER TABLE t STATS_OPTIONS = 'str' | DEFAULT` (pkg/parser/parser.y:2283)
- AST `ast.StatsOptionsSpec` and `ast.AlterTableStatsOptions` (pkg/parser/ast/ddl.go:5830, :4056)
- Metadata `model.TableInfo.StatsOptions *model.StatsOptions` (pkg/meta/model/table.go:210) holding
AutoRecalc, ColumnChoice, ColumnList, SampleNum, SampleRate, Buckets, TopN, Concurrency and an
embedded `*StatsWindowSettings`
- DDL action `model.ActionAlterTableStatsOptions = 58` (pkg/meta/model/job.go:102), referenced only
by BDR role tests

There is no executor: `ast.AlterTableStatsOptions` is absent from the ALTER TABLE switch in
pkg/ddl/executor.go, so the statement falls through to `default: ErrUnsupportedAlterTableSpec`, and
nothing anywhere reads `TableInfo.StatsOptions`. Because the JSON field has existed for years and
was never written, reviving it needs no TableInfo version bump and old binaries simply ignore it.

Note also that MySQL's `STATS_PERSISTENT`, `STATS_AUTO_RECALC` and `STATS_SAMPLE_PAGES` table
options are already parsed and ignored by TiDB (pkg/parser/misc.go:808-819), so any new option
names must not collide with them and should not be confused with them in documentation.

**Describe alternatives you've considered:**

1. Keep `mysql.analyze_options` as the single store and have CREATE/ALTER TABLE write rows into it.
Smallest change and one source of truth, but the write is not atomic with the schema change, is
invisible to `SHOW CREATE TABLE`, and keeps the truncate / exchange partition / BR data loss.
2. Store in `TableInfo`, finishing the scaffolding above. Survives truncate, exchange partition and
BR restore, appears in `SHOW CREATE TABLE`, versioned like any other schema change. Cost: two
stores, so an explicit precedence rule against `mysql.analyze_options` is required.
3. A new system table keyed by schema and table name instead of id. Fixes truncate and BR without
DDL work, but invents a third stats configuration store and still misses `SHOW CREATE TABLE`.

**Teachability, Documentation, Adoption, Migration Strategy:**

Proposed precedence, most specific first:

statement WITH options
> options persisted by a previous ANALYZE ... WITH (mysql.analyze_options)
> options declared in DDL (TableInfo)
> tidb_analyze_default_num_buckets / _num_topn and the other system defaults

`DEFAULT` is the reset spelling at every level, consistent with #69853.

Scope: table level first. Per-index and per-column options do not exist anywhere today, since
`mysql.analyze_options` is keyed by physical table id and only `column_choice` / `column_ids` come
close, so they need their own design and should be a follow-up.

**Open questions for discussion:**

1. Store: `TableInfo` (alternative 2) or `mysql.analyze_options` (alternative 1)?
2. Syntax: first-class options (`STATS_BUCKETS`, `STATS_TOPN`, ...) or the existing
`STATS_OPTIONS='buckets=64,topn=20'` string blob already in the grammar?
3. Does `tidb_persist_analyze_options = OFF` suppress DDL-declared options, or are they schema and
therefore always in effect?
4. Partition level: is `ALTER TABLE t PARTITION p STATS_...` in scope? This interacts with #69952.
5. What happens to the unused `StatsWindowSettings` and `Concurrency` fields on `model.StatsOptions`:
keep unset, or strip them in the same change?

Start by reading the existing grammar in pkg/parser/parser.y, the AST types in pkg/parser/ast/ddl.go, and the ALTER TABLE dispatch in pkg/ddl/executor.go. Trace model.TableInfo.StatsOptions in pkg/meta/model/table.go and ActionAlterTableStatsOptions in pkg/meta/model/job.go, then resolve the open questions about syntax, storage, precedence, and partition scope. Done means the chosen design is implemented for table-level CREATE and ALTER TABLE options, with DEFAULT clearing values and no statistics collection.

Written by the indexing model from the issue text.