The scalar_at hot path was creating and dropping an ExecutionCtx on
every call via LEGACY_SESSION.create_execution_ctx(). Each creation
involves an atomic fetch_add + Arc clone, and each drop an Arc release.
This was called millions of times during query execution (primarily from
PrimitiveTyped::value_unchecked during search_sorted and patches).

Only Patched::scalar_at actually uses ctx (for .execute()), so move the
ctx creation there. All other implementations had unused `_ctx` params.

Profiled with apmc (Apple Silicon hardware performance counters).
ClickBench (Vortex format, Apple Silicon M4 Max, 3 iterations):

  Metric               Before          After           Delta
  Cycles               1,973.5B        1,942.6B        -1.6%
  Instructions         4,730.2B        4,722.5B        -0.16%
  IPC                  2.40            2.43            +1.25%
  Wall clock           43.09s          41.92s          -2.7%
  Dispatch stalls      593.7B          573.9B          -3.3%
  L1D cache misses     80.7B           80.6B           ~0%
  Branch mispredicts   11.2B           11.2B           ~0%

Signed-off-by: Alexander Droste <droste.alexander@gmail.com>

Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>