Create Nested(...) columns as Array(Tuple(...))

[Hakaou]

Closes #571.

Summary

TypeParser had no entry for Nested, so the name became a Void terminal. At the root ValidateAST rejects it ("unsupported column type"), but inside a container nothing checks it: Array(Nested(key String, value String)) was created as Array(Void), ColumnNothing reads one byte per row and the block stream desynchronises.

Nested is now its own AST node, next to Map and SimpleAggregateFunction, and the column factory builds it in the Tuple case, wrapped in an Array. That is how the server defines it: DataTypeNested.cpp builds DataTypeArray(DataTypeTuple(types, names)) under a custom name. Sharing the Tuple case keeps the AST read by reference, so a deeply nested type is not copied at every level, the pattern #588 removes from LowCardinality.

Behavior changes

• Array(Nested(...)) and Map(K, Nested(...)) columns decode, with their field names kept on the Tuple.
• A top-level Nested column, which the server sends with flatten_nested = 0, decodes instead of failing with "unsupported column type".
• The created column reports the structural type name Array(Tuple(key String, value UInt8)) rather than Nested(...). The server accepts that name on INSERT: a block selected from such a column goes back in as is (ArrayOfNested_ReinsertSelectedRows), and clickhouse local 26.11 also accepts unnamed tuples. A dedicated Nested column type would add public API for no behavioural gain, so I did not go that way. Happy to discuss.
• TypeAst::Meta gains the Nested value.

Tests

• type_parser_ut: the Nested node with its named elements, and a field itself named Nested.
• CreateColumnByType_ut: Nested, Array(Nested), Map(String, Nested) and Nested(a Nested(...)) give the expected column types. DeepNestedDoesNotCopyAst builds 200 levels of Nested over 8 000 fields and checks that peak RSS grows by less than 16 MiB, about 180 MiB when the node is copied at each level. The measure runs in a child process (EXPECT_EXIT, hence the DeathTest suite) with its own peak counter, so a peak reached earlier in the test binary cannot hide a regression. It reads getrusage, so it runs on Linux and macOS and skips on Windows.
• client_ut, both fixtures (no compression and LZ4): select from a server-typed Array(Nested(...)) column with empty inner and outer arrays, insert a selected block back as is and read it back with toString, and a top-level Nested column with flatten_nested = 0.

Each new test fails when the Nested entry in GetTypeMeta is removed. Before this change, on master fc5d98f:

Case	Before	After
CreateColumnByType("Array(Nested(key String, value UInt8))")	Array(Void)	Array(Array(Tuple(key String, value UInt8)))
Select on that column, no compression	Can't cast from Void	3 rows, values match
Select on that column, LZ4	some data was not read	3 rows, values match
Select on a Nested column, flatten_nested = 0	unsupported column type: Nested(key String, value UInt8)	2 rows, values match
Select, then insert the block back	fails at the Select	both rows read back identically

Full suite green locally (macOS arm64, Apple clang 21) against ClickHouse 26.11.1 master build, the new client tests also against 25.12 and 26.3. Under AddressSanitizer and UBSan, the parser and factory suites pass on this branch, and the client and roundtrip suites with #590 on top.

Out of scope

• ValidateAST only checks the root, so any other unknown type nested in a container still decodes as Void silently, the second point of the suggested fix in Array(Nested(...)) columns silently decode as Array(Nothing) and desynchronize the native block stream #571. Tracked separately in Unknown type names inside a container decode silently as wrong data #589, with a draft in Reject unknown type names inside containers #590.
• The existing InsertData test reads past TEST_DATA under UBSan and depends on block order. Test code only, fixed separately in Make InsertData independent of block boundaries and order #576.

[CLAassistant]

All committers have signed the CLA.

[Hakaou]

Nested takes the Tuple path and is wrapped here, so its elements are read from the AST by reference. An earlier version of this PR copied the node into a Tuple first, which costs the subtree at every level: about 180 MiB of peak RSS for 200 levels over 8 000 fields, the pattern #588 removes from LowCardinality. DeepNestedDoesNotCopyAst holds it.