Fix #33425: Handle F16/BF16 constants in If operator constant folding

[vishalharkal15]

This commit fixes a compiler crash that occurred when compiling models with If operators containing F16 or BF16 constants during constant folding.

Problem:

• When compile_model() performs constant folding on If/Loop operators, it evaluates subgraphs to compute constant outputs
• F16/BF16 tensors were being accessed with F32 pointers, causing: 'Check is_pointer_representable(element_type) failed at src/inference/src/dev/make_tensor.cpp:83'

Root Cause:

• The function() helper in src/core/reference/src/op/function.cpp created output tensors matching the result type directly
• F16/BF16 types cannot be safely accessed as F32 pointers
• This led to a type safety violation during evaluation

Solution:

• Detect F16/BF16 result types before creating output tensors
• Create F32 tensors for evaluation (compatible with pointer access)
• Evaluate the subgraph with F32 tensors
• Convert results back to F16/BF16 after evaluation
• Both function() overloads updated for consistency

Testing:

• Comprehensive test suite added in test_if_f16_fix.py
• Tests cover: ONNX model loading, compilation, inference, and programmatic model creation with F16 constants
• Validates both then/else branches of If operator

Fixes: #33425

Details:

• item1
• ...

Tickets:

• ticket-id