Vector Verification: Rework to run based on env variable `DUCKDB_DEBUG_VERIFY_VECTOR` and move to `Main.yml` (duckdb/duckdb#16957)

Vector Verification: Rework to run based on env variable `DUCKDB_DEBUG_VERIFY_VECTOR` and move to `Main.yml` (duckdb/duckdb#16957)

Vector Verification: Rework to run based on env variable `DUCKDB_DEBUG_VERIFY_VECTOR` and move to `Main.yml` (duckdb/duckdb#16957)

Vector Verification: Rework to run based on env variable `DUCKDB_DEBUG_VERIFY_VECTOR` and move to `Main.yml` (duckdb/duckdb#16957)

Vector Verification: Rework to run based on env variable `DUCKDB_DEBUG_VERIFY_VECTOR` and move to `Main.yml` (duckdb/duckdb#16957)

#16957 fixed the bug in `Vector::Reinterpret` for dictionary vectors by
creating a new child buffer every time. However, creating a new child
buffer is much more expensive than copying a shared_ptr, which results
in about 3% performance regression in TPC-H benchmarks.

The main issue is that a dictionary vector cannot directly copy the
child buffer if the types don't match, as this would make the child
vector's type inconsistent. So we can check whether the types are
different and only create a new child buffer when the types are not
equal - which isn't the case in most scenarios.

benchmark result:
```
python scripts/regression/test_runner.py --old build/release-old/benchmark/benchmark_runner --new build/release/benchmark/benchmark_runner --benchmarks .github/regression/tpch.csv --threads 2
Old timing geometric mean: 0.02501917589777838
New timing geometric mean: 0.024095980407583723, roughly 3% faster

python scripts/regression/test_runner.py --old build/release-old/benchmark/benchmark_runner --new build/release/benchmark/benchmark_runner --benchmarks .github/regression/tpcds.csv --threads 2
Old timing geometric mean: 0.021899101352704627
New timing geometric mean: 0.021254155592682008, roughly 2% faster
```