Optimize 2D tensor gathering to skip sensitive layers early

[gacty]

Summary

Adds early filtering for normalization/modulation layers at the 2D tensor gathering stage (line 261)

Problem

Currently, sensitive layers (norms, modulations, embeddings) are gathered into weight_keys and then filtered out later in the exclusion logic (around line 324+). This means they're loaded into memory unnecessarily

Solution

Move common exclusion patterns up to the gathering stage using the existing AVOID_KEY_NAMES constant:

and not any(avoid in key for avoid in AVOID_KEY_NAMES)

This acts as a "first line of defense" before the MODEL_FILTERS logic

Benefits

Reduces the memory footprint of weight_keys list
Will avoid unnecessary downstream filtering for already-excluded layers
Complements an existing MODEL_FILTERS system
No behavioral changes