[Feature] GLA CPU reference implementation and comparison tests

[0xaskr]

Motivation

Part of Epic #28. GLA (Gated Linear Attention) is the highest priority flash-linear-attention component requiring a CPU JAX gold reference for kernel correctness validation.

Expected Behavior

• Complete CPU JAX implementation of GLA in tops/cpu/ops/gla/
• Supports the GLA recurrence: h_t = h_{t-1} * exp(gk_t) + k_t^T @ v_t, o_t = q_t^T @ h_t
• Input layout follows project convention: [B, T, H, K]
• Comprehensive docstrings with tensor shape and dimension semantics
• Strict input assertions on shape and types
• Comparison tests in tests/ops/gla/ validating Pallas kernels against CPU reference
• Tolerance-based assertions using compare_tensor(name, gold, tensor, atol, rtol, max_ulp)

Proposed Approach

1. Implement CPU reference in tops/cpu/ops/gla/ covering fused recurrent and chunk variants
2. Add comparison tests in tests/ops/gla/
3. Error bound constraint: TPU error ≤ GPU error
4. Export public API via tops/ops/__init__.py

Willingness to Contribute

• I am willing to submit a PR for this feature.

Additional Context

Sub-task of #28.