[Feature] MLA CPU reference implementation and comparison tests

[0xaskr]

Motivation

Part of Epic #28. MLA (Multi-head Latent Attention) is one of the three priority flash-linear-attention components requiring a CPU JAX gold reference for kernel correctness validation.

Expected Behavior

• Complete CPU JAX implementation of MLA in tops/cpu/ops/mla/
• 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/mla/ 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/mla/
2. Follow existing GLA reference implementation patterns
3. Add comparison tests in tests/ops/mla/
4. Error bound constraint: TPU error ≤ GPU error
5. 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.