[CPU] Stop scaling PackOp distribution tiles by inner pack sizes (#24107)

Undo a 3-year old tiling tweak (#11873) that should not be necessary,
though if it ever helped performance, that would indicate suboptimal
codegen.

The rationale, stated in a comment, is that as packing is an
element-wise operation, the only reason to need multiple tiles is to
distribute work to multiple threads. Memory locality should not be a
concern; if it were a concern, that would be a codegen bug (suboptimal
traversal within the tile).


[Benchmark](https://docs.google.com/spreadsheets/d/1q0pY8uAI4ApGSUndJYAdSq7CKGRzTJgaWlEzMkvSL9A/edit?usp=sharing)
(3 PRs together: #24107, #24108, #24109) showing no regression across
many matmul shapes and slight improvement to scaling SDXL-clip to many
threads.

The point is not the tiny perf improvement per se, but ensuring that
existing parameters have the intended semantics.

Made-with: Cursor

Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>