Implement binary search for bfp4 weight lowering based on sensitivity

[kdimicTT]

Problem description

When running models with mixed precision, not all weights tolerate aggressive quantization equally. Lowering every weight to bfp_bf4 maximizes memory savings but often degrades accuracy below acceptable thresholds. There was no systematic way to find the optimal subset of weights to lower to bfp_bf4 while keeping accuracy above a target.

The sensitivity scores consumed by this script are produced by the per-weight scoring implemented in #5092.

What's changed

Adds mixed_precision/lowering_search.py script that automates the search for the maximum number of weights that can be safely lowered to bfp_bf4.

Approach:

1. Reads a pre-computed sensitivity scores JSON (weights ranked from most to least sensitive) for the target model.
2. Runs three reference baselines: all-bfp_bf8, all-MLP-at-bfp_bf4, all-bfp_bf4.
3. Binary searches over the count k of least-sensitive weights to lower, evaluating each candidate by running the accuracy benchmark test and parsing the TOP1 p5 metric against a threshold.
4. Writes the resulting mixed-precision config JSON and a markdown results report with optional per-iteration logs.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 33.84%. Comparing base (8f71001) to head (7f977f4).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5418   +/-   ##
=======================================
  Coverage   33.84%   33.84%           
=======================================
  Files          37       37           
  Lines        4990     4990           
=======================================
  Hits         1689     1689           
  Misses       3301     3301           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.