[CPU] Block root op chain fusion in non-root anchored tiling. (#24098)

When tiling from a non-root anchor op at VectorInnerParallelTiles level,
there are two paths, because the intention is for tiling dimensions that
not captured by rootOp:
(a) Start from an anchor op before the root op.
(b) Start from an anchor op after the root op.

In both cases, we should not fuse rootOp and the consumers/producers for
(a)/(b).

The revision aims to fix the case that reduction dimension is not tiled
on the root op which fused the root op unintended; it results in numeric
issues as expected. This can happen when the reduction dimension is much
less than native vector sizes.

Signed-off-by: hanhanW <hanhan0912@gmail.com>