Represent scan snapshots explicitly instead of parsing table names

[ila]

Thanks for reviewing.

Why

LPTS currently appends AT (...) to a table-name string and later reparses that string to render a snapshot. This confuses identifiers with SQL: an ordinary table named events AT (VERSION => 2) is incorrectly treated as a time-travel scan.

OpenIVM also needs to restore foreign snapshot metadata after binding against local schema-only tables. Doing that in a separate AST walk duplicates traversal and relies on the same string convention.

Change

• Carry a parsed AtClause separately on scan AST/CTE nodes; render it in DuckDB or Spark syntax at serialization time.
• Remove the table-name suffix splitting/quoting helpers.
• Add an optional snapshot resolver to LogicalPlanToAst, invoked during its existing GET visit with the original catalog entry, before output-name remapping affects identity. Native DuckLake snapshot metadata takes precedence.
• Extend the existing time-travel and Spark dialect tests for escaped identifiers, alias column lists, merged/unmerged rendering, unqualified output, and identifiers that resemble snapshot syntax.

Production code is 49 lines smaller; tests add 78 lines. No new settings, additional plan walk, or handwritten SQL parser.

Validation

• LPTS time-travel tests: 83 assertions passed.
• Spark dialect tests: 69 assertions passed, including actual lpts_check bag-equality validation of the quoted-name regression.
• Full LPTS SQL suite: 2,352 assertions across 46 cases passed; the initially skipped TPC-H dependency was then installed and that case separately passed all 24 assertions (all 47 cases exercised).
• make test DuckDB corpus gate: all 3,346 files completed, no new WRONG or FAIL against baseline.
• Consumer integration in OpenIVM PR Openivm postgres implementation #10: time-travel suite passed 214 assertions, including batched mixed DML followed by refresh and bidirectional EXCEPT ALL. Broader OpenIVM suite passed 10,577 assertions across 88 cases; one ICU-dependent case was skipped by that runner.
• SQLStorm SF0.001 is still running. A baseline comparison is being prepared to distinguish existing errors/timeouts from regressions. Keeping this draft until that check and final review finish.

The OpenIVM consumer update is being prepared separately on ila/openivm#10; it removes the duplicate input scanner and uses this resolver during the existing LPTS traversal. Some text-only OpenIVM refresh paths still retain their existing restoration scanner; this LPTS change does not claim to eliminate those paths.