Add comprehensive unit and GPU parity tests for core ops

[rlarson20]

Summary

This PR significantly expands test coverage for the Volta tensor library by adding ~1000 lines of direct unit tests for foundational operations and neural network layers, plus a new GPU parity test suite to validate CPU/GPU numerical equivalence.

Key Changes

Unit Tests (in src/lib.rs)

• Binary ops tests (binary_ops_tests): Forward value/shape and gradient checks for add, sub, elem_mul, div, max_elem, modulo, cmplt, and chained operations
• Unary ops tests (unary_ops_tests): Forward correctness and gradient validation for 15+ operations including neg, recip, sqrt, exp, log, sin, cos, tanh, sigmoid, relu, erf, and exponential/logarithmic variants
• Reduction ops tests (reduce_ops_tests): Tests for sum, mean, max_reduce and their axis-aware variants (sum_dim, mean_dim, max_dim) with shape and gradient correctness
• Matmul tests (matmul_tests): Coverage for 2D×2D, 2D×1D, 1D×2D, 1D×1D (dot product), and batched matmul with forward value and gradient checks; includes transpose validation
• Linear layer tests (linear_tests): Weight/bias shape validation, forward computation with known weights, and gradient checks
• Activation layer tests (activation_tests): Forward and backward parity for ReLU, Sigmoid, and Tanh modules
• Dropout tests (dropout_tests): Identity behavior at p=0, zeroing at p=1, eval-mode passthrough, shape preservation, and statistical mean preservation
• BatchNorm tests (batchnorm_tests): Normalization correctness, running statistics updates, eval-mode behavior, and gradient flow for both BatchNorm1d and BatchNorm2d

GPU Parity Test Suite (new files)

• tests/common/mod.rs: Shared helpers for CPU/GPU parity testing with configurable tolerances and early exit when GPU is unavailable
• tests/parity_unary.rs: CPU/GPU forward+backward parity for 15+ unary operations
• tests/parity_binary.rs: CPU/GPU parity for binary ops (add, sub, mul, div, max_elem, modulo, cmplt)
• tests/parity_reduce.rs: CPU/GPU forward parity for reductions (sum, mean, max with axis variants)
• tests/parity_matmul.rs: CPU/GPU parity across the matmul shape matrix (2D×2D, 2D×1D, 1D×2D, 1D×1D, batched)

Implementation Details

• All unit tests use RawTensor::check_gradients_simple for finite-difference gradient validation, ensuring backward-pass correctness
• GPU parity tests use deterministic input generation with configurable value ranges to avoid degenerate cases (e.g., log of zero, division by zero)
• Tests gracefully skip when GPU is unavailable (skip_if_no_gpu() guard), allowing compilation and CI to pass on CPU-only environments
• Tolerances are tuned per operation class (forward: 1e-5, backward: 1e-4 for most ops; matmul: 1e-4/1e-3 due to accumulation)
• Tests follow existing inline #[cfg(test)] mod convention and use approx::assert_relative_eq for floating-point comparisons where appropriate

https://claude.ai/code/session_01BPobYZJo4z4VbKFpVFRNpj