[CPU] Fix tile sizes based on number of threads. (#24108)

The requested number of threads T was effectively used as a divisor for
the tile size along each dimension separately. This effectively resulted
in targeting T^N threads, where N is the number of tile dimensions.

This PR it to instead be a divisor on the overall number of elements.


[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>