NPUW: Fix cases when FOLDing is not enabled

[dmatveev]

Details:

• CompiledModel: do API-modifying transformations before passing the model down to ICompiledModel constructor

  • Without that, compiling model with weights and SLICE_OUT:YES caused a mismatch in allocated/set tensor size. NPUW's CompiledModel first initializes its inputs() / outputs() via the parent class, ov::ICompiledModel, and then modifies the shape of the output tensor if SLICE_OUT transformation worked.
  • When an Infer Request is created, it's outputs are allocated based either on the model information (above) or the function information, if produced by a function.
  • SLICE_OUT altered function bodies so when the last graph's output was a function (WITH folding), the meta mismatch was ignored. When folding is disabled, however, the output tensor is created based on the original (untransformed) model information and then is passed to a transformed subgraph, causing an assert failure.
  • Moving the parent-model-side API-altering transformations before initializing the NPUW's CompiledModel solved the problem.

• UnfoldInferRequest: run unpack_closure for functions only.

  • If FOLDing is disabled, there's no functions but unpack_closure will still be called, leading to an assert.

Tickets:

• ticket-id