fix(embedding): preserve retained casts and protect destructive MO inputs

[aunjgr]

Public MatrixOne TPC-H execution exposed two GPU correctness failures: Q7 failed on a retained VARCHAR identity cast, and Q15 could silently split a destructive MO input between consumers of a shared CTE.

Exact Substrait preparation skips DuckDB's ordinary optimizer and retains semantic VARCHAR → VARCHAR casts. Preserve string values and validity by transferring an owned column or copying a borrowed view on the caller's stream and memory resource, instead of passing STRING to cuDF's fixed-width cast.

Validate reachable read consumption through ReferenceRel before admitting work. Shared references cannot replay a destructive MO input, and unreachable MO bindings cannot start producers that never receive a consumer. A single reference and immutable TAE replay remain valid. Reference expansion has explicit node/depth bounds. MatrixOne's accompanying fix gives each inlined MO occurrence an independent reader on the same snapshot; canonical SQL stays unchanged.

Update the value-variant test to guard the already merged, appended exact coefficient at index 21. Its stale 21-alternative assertion prevented the owning test target from compiling. No native API or numeric ABI layout change.

Validation:

• VARCHAR ownership, empty/ASCII/Unicode values and NULL under MATERIALIZE, AST_INTERPRET and AST_JIT: 22 assertions passed. Owning expression suite: 1,383 assertions in 18 cases passed.
• CUDA memcheck over cast/value cases: 106 assertions in 28 cases, zero errors.
• Shared-reference regression fails on the previous validator because it accepts two consumers of one MO binding; passes with this fix. Full binding suite: 856 assertions in 82 cases passed.
• Current control suite: 606 assertions in 78 cases passed. Freshly relinked GPU consumers: 30,635 assertions in 18 cases passed.
• Public MatrixOne development validation: canonical Q7 and Q15 match native MO, including complete rows/descriptors, completed GPU work, healthy cleanup and zero retained input/result credit. All 22 SF1 queries have passing smoke evidence. These checks are development evidence, not full campaign acceptance.
• Changed-file hooks passed.

Prerequisite for MatrixOne #28968. Base: upstream-dev-merge.