fix(mssql): Generate valid SQL for withRecursive() [#4514]

[jeremy-w]

mssql forbids use of the RECURSIVE keyword required by some other
dialects.

The approach taken here maximizes subclass reuse and leaking
changes to the query outside with generation. If we skipped
restoring the recursive labels using undoList, the SQL would
generate fine here, but the test suite fails:
The next-tested driver fails its test, because clone() preserves the
top-level statement's recursive = false change, so it's shy one
RECURSIVE in the generated SQL.

This addresses the mssql portion of [#4514], but not the oracledb issues.
If this approach to omitting RECURSIVE looks good, I can apply the same to the
oracledb dialect. (Supporting a column alias list as separate argument, as Oracle
requires for recursive CTEs, would require deeper changes.)

[jeremy-w]

first pass at integration test had errors on a couple platforms i wasn't testing locally:

• postgres: the format of returning must vary across drivers - pg was returning [{id: 1}] rather than just [1]
• mysql: its increments() column is apparently unsigned, and so it refused to let a (signed) integer column reference the id column.

i revised it to omit the keys and fkeys (they weren't the point of the test - just superfluous "realism"), and everything passed the next time around. 👍

[kibertoad]

Thanks a lot! Could you also add same for Oracle?

[jeremy-w]

Started poking at it - got Oracle working locally so I can iterate on it.

Oracle requires a column list to be valid (with “cteName” (“col1”, “col2”) as …), so it might take larger changes, unless using a raw value in the first argument works as a workaround.

But while trying to check that workaround, I somehow got the query compiler to blow the stack with recursion - haven’t had a chance to untangle that unexpected diversion just yet.

[jeremy-w]

Recursion was due to code like this.select().unionAll(this.select()). wrapping the unionAll arg in a callback resolved it.

[jeremy-w]

Opened #4652 to apply the parallel change made here. We'll need to work out a querybuilder API change to be able to successfully issue rCTE queries on Oracle, though, in order to accommodate its requirement of a column list.

[kibertoad]

Released in 0.95.11