torch inference for segmentation

[alisterburt]

this PR is on top of #3 and adds

• a torch implementation of the segmentation model
• some utilities for converting weights
• some code to validate that implementations match numerically

This was heavily LLM assisted, generally okay but had to work around some differences in padding behavior between TF and torch

Not really expecting this to be merged directly, is just good to know it's possible to run inference using your weights in torch

============================================================
PyTorch vs TensorFlow Model Equivalence Validation
============================================================

============================================================
WEIGHT VALIDATION
============================================================
TensorFlow parameters:
  Trainable: 128,897,505
  Total: 128,917,537

PyTorch parameters:
  Trainable: 128,897,505
  Total: 128,897,505

✓ PASS: Parameter counts match

============================================================
MULTIPLE INPUT VALIDATION (5 tests)
============================================================

Test 1/5:

============================================================
FORWARD PASS VALIDATION
============================================================
Test input shape (TF format): (64, 64, 64, 1)
TensorFlow output shape: (1, 64, 64, 64, 1)
TensorFlow output range: [0.000000, 0.999970]
PyTorch output shape: (1, 64, 64, 64, 1)
PyTorch output range: [0.000000, 0.999970]

Numerical Differences:
  Max absolute difference:  2.35e-06
  Mean absolute difference: 7.40e-08
  Max relative difference:  1.28e-05
  Mean relative difference: 9.34e-07

✓ PASS: Outputs match within tolerance (rtol=0.0001, atol=1e-05)

Test 2/5:

============================================================
FORWARD PASS VALIDATION
============================================================
Test input shape (TF format): (128, 128, 128, 1)
TensorFlow output shape: (1, 128, 128, 128, 1)
TensorFlow output range: [0.000000, 0.999988]
PyTorch output shape: (1, 128, 128, 128, 1)
PyTorch output range: [0.000000, 0.999988]

Numerical Differences:
  Max absolute difference:  2.32e-06
  Mean absolute difference: 3.67e-08
  Max relative difference:  1.72e-05
  Mean relative difference: 1.07e-06

✓ PASS: Outputs match within tolerance (rtol=0.0001, atol=1e-05)

Test 3/5:

============================================================
FORWARD PASS VALIDATION
============================================================
Test input shape (TF format): (64, 64, 64, 1)
TensorFlow output shape: (1, 64, 64, 64, 1)
TensorFlow output range: [0.000000, 0.999902]
PyTorch output shape: (1, 64, 64, 64, 1)
PyTorch output range: [0.000000, 0.999902]

Numerical Differences:
  Max absolute difference:  1.52e-06
  Mean absolute difference: 7.58e-08
  Max relative difference:  1.43e-05
  Mean relative difference: 9.38e-07

✓ PASS: Outputs match within tolerance (rtol=0.0001, atol=1e-05)

Test 4/5:

============================================================
FORWARD PASS VALIDATION
============================================================
Test input shape (TF format): (128, 128, 128, 1)
TensorFlow output shape: (1, 128, 128, 128, 1)
TensorFlow output range: [0.000000, 0.999987]
PyTorch output shape: (1, 128, 128, 128, 1)
PyTorch output range: [0.000000, 0.999987]

Numerical Differences:
  Max absolute difference:  2.74e-06
  Mean absolute difference: 3.74e-08
  Max relative difference:  1.70e-05
  Mean relative difference: 1.09e-06

✓ PASS: Outputs match within tolerance (rtol=0.0001, atol=1e-05)

Test 5/5:

============================================================
FORWARD PASS VALIDATION
============================================================
Test input shape (TF format): (64, 64, 64, 1)
TensorFlow output shape: (1, 64, 64, 64, 1)
TensorFlow output range: [0.000000, 0.999940]
PyTorch output shape: (1, 64, 64, 64, 1)
PyTorch output range: [0.000000, 0.999940]

Numerical Differences:
  Max absolute difference:  2.62e-06
  Mean absolute difference: 7.60e-08
  Max relative difference:  2.57e-05
  Mean relative difference: 1.06e-06

✓ PASS: Outputs match within tolerance (rtol=0.0001, atol=1e-05)

✓ ALL TESTS PASSED (5/5)

============================================================
VALIDATION SUMMARY
============================================================
Weight validation:  ✓ PASSED
Forward pass tests: ✓ PASSED

🎉 ALL VALIDATIONS PASSED!

[alisterburt]

ooft - diff is ugly, code changes from #3 are in src/easymode/segmentation/torch_version