Dim Order Validation Inconsistency for Edge / Ambiguous Cases

[psiddh]

tensors_have_same_dim_order() fails incorrectly for memory-compatible tensors with ambiguous dimensions

Problem Summary
ExecuTorch's dim order validation has an inconsistency where tensor_is_contiguous() and tensors_have_same_dim_order() use different validation approaches, causing false negatives for edge cases with ambiguous dimensions (C=1 or H=W=1).

Root Cause Analysis

1. tensor_is_contiguous() - Memory layout validation ✅

• Checks actual stride patterns to determine if memory is contiguous
• Correctly handles edge cases where different dim_order labels produce identical memory

2. tensors_have_same_dim_order() - Label-based validation ❌

• Checks if all tensors have identical dim_order labels (strict matching)
• Fails when tensors have compatible memory but different labels

3. The Inconsistency

• Same tensor data can pass individual contiguous checks but fail cross-tensor validation
• Edge cases with ambiguous dimensions are incorrectly rejected

When This Happens
Ambiguous Dimension Cases:
C=1: Channel dimension = 1 makes NCHW/NHWC memory layouts identical
H=W=1: Spatial dimensions = 1 makes NCHW/NHWC memory layouts identical
Multiple dims = 1: Combined edge cases

========================

// Both tensors have IDENTICAL memory layout
Tensor nchw = tf_float_.ones({2, 1, 4, 4}); // dim_order: [0,1,2,3]
Tensor nhwc = tf_float_.full_channels_last({2, 1, 4, 4}, 1.0); // dim_order: [0,2,3,1]
// Memory layout: N₀N₁ H₀W₀W₁W₂W₃ H₁W₀W₁W₂W₃ H₂W₀W₁W₂W₃ H₃W₀W₁W₂W₃
// ↑ IDENTICAL for both tensors because C=1
// Individual validation:
tensor_is_contiguous(nchw) == true ✅
tensor_is_contiguous(nhwc) == true ✅
// Cross validation:
tensors_have_same_dim_order(nchw, nhwc) == false ❌ BUG!

========================

Also note that the existing unit test cases do not seem to cover these edge cases