refactor: optimized covariance transform in ExpectedAttentionPress

[neuralsorcerer]

Changes:

• Compute per-head query covariance directly in the projected query space, avoiding any intermediate $O((n * d)^2)$ hidden-state covariance tensors.

Why?

• A quick benchmark with n=32, d=128, seq_len=64 showed the old method taking ~1.87s and storing 33,554,432 elements vs. ~0.04s and 1,048,576 elements for this approach about 32x less memory and ~50x faster.

[alessiodevoto]

Hi @neuralsorcerer ! I just revised and tested the code and I believe we can merge, thanks for opening this PR and contributing to KVPress !