[CPU] fix CPU FullyConnected Node zeros value output issue

[liubo-intel]

Details:

• CPU internal FullyConnected Node will use cached constant weights during inference, while this will make the output of this node always zero values when its weight 'constant_path' conatains 'RandomUniform' Operation. so need to change current 'constant_path' logic to keep use Matmul kernels for these cases.

Tickets:

• CVS-165978

[liubo-intel]

Hi, @luweizhou2016 : could you please help review this pr when you are free? thanks.

[luweizhou2016]

It should be a bug on is_on_constant_path(). why not try more general fixing? The RandomUniform() node can't be const folded.

Suggested change

	if (current_node->get_input_size() == 0 && !ov::is_type<ov::op::v0::Constant>(current_node)) {
	if ((current_node->get_input_size() == 0 && !ov::is_type<ov::op::v0::Constant>(current_node)) ||
	current_node->can_constant_fold(current_node->input_values())){

[liubo-intel]

Hi, @luweizhou2016 : do you mean with this changes || !current_node->can_constant_fold(current_node->input_values()) ?
but it seems with this kind of changes, we will limit the condition to constant inputs only, while this seems will make current 'constant_path' logic miss some cases, e.g. this specific following kind( the following second 'Eltwise' pattern should be suitable pattern for 'constant_path' logic if none the 'RandomUniform'):

[luweizhou2016]

OK. seems can_constant_fold() can't propagate to the graph input.
What about " || ov::is_type(ov::op::v8::RandomUniform)", the key is this should be a bug and should be fixed in the is_on_constant_path() internally not by the API user. Once the graph path has the RandomUniform OP, it should return false.