Fix stale bounds after AbsSinking via conditional invalidation in replace_output()

[evkotov]

Details:

Problem
VIT model compilation fails with "Value for partial shape evaluation can't be lower than zero" because:

• SymbolicPropagation sets SkipInvalidation attribute on tensors to prevent bounds recalculation during symbolic optimization
• AbsSinking transformation modifies Concat inputs (inserts Abs before Concat)
• Concat bounds are not invalidated due to SkipInvalidation, keeping stale values with -1
• Broadcast shape inference receives stale bounds with -1 => assertion failure

Solution
Implement conditional bounds invalidation in replace_output():

• Invalidate bounds when old source has bounds but new source doesn't (AbsSinking case - Abs output has no pre-computed bounds)
• Invalidate bounds when both have bounds but they differ
• Do not invalidate when bounds are the same (preserves OptimizeSymbolsUsedAsValues optimization)
• New helper functions in bound_evaluation_util.hpp:
  tensors_equal() - compares two tensors for equality
  force_invalidate_bounds() - invalidates bounds bypassing SkipInvalidation attribute

Tickets:

• 175062

[praasz]

is not duplicate of are_equal ?
This helper should do same.

[evkotov]

No, it's not a duplicate. The key difference is in handling empty tensors

[praasz]

are_equal also handles empty tensors, most of it is duplication.

[praasz]

These check looks like duplicate whe are_equal do.
Do we need check tensors or just check in new input and all input are not same tensors (without data check)?

[evkotov]

In OptimizeSymbolsUsedAsValues, we call evaluate_both_bounds() and then replace(). The old and new outputs point to different nodes but may have identical bound values (same shape data). Using are_same_tensor (pointer comparison via data()) would return false and trigger unnecessary invalidation.

[evkotov]

Created #33622 with alternative approach