Fix float32 MLE optimization with autograd gradients

[AparajeetS]

Fixes #61.

What changed

• compute the convex MLE objective gradient with PyTorch autograd
• pass the exact Jacobian to scipy.optimize.minimize
• use the same Jacobian path for masked re-optimization
• add a deterministic regression test for TemperatureScaling and
  LogisticCalibration with both float32 and float64 inputs
• verify that the repaired float32 loss agrees with the float64 control

Root cause

The MLE callback converted SciPy's float64 parameter vector to the prepared input
tensor's dtype. With float32 inputs, SciPy's default finite-difference perturbation near
one was rounded away by that conversion. The loss at w=1 and w=1+1e-8 became
identical, so SciPy reported a zero gradient after two evaluations and returned the
identity initialization with a successful status.

On a deterministic five-class overconfidence problem, the old float32 behavior kept the
weight at 1.0 and NLL at 1.8382. The autograd path fits a weight of 0.3286 and
reduces NLL to 1.3154, matching the float64 control (0.3285, 1.3154).

Validation

• python -m unittest tests.test_mle_precision -v
• direct patched-library reproduction for float32 and float64
• python -m compileall -q netcal tests
• git diff --check

The full executable diagnosis and recorded results are available in
Phason Evidence Audit PEA-005.