fix(sparse_probing): add std_floor to standardization

[VictorAraujopy]

Fixes #1816

Under preprocess="standardize", only an exactly-zero std was guarded, so a near-constant column was divided by its own tiny std and reached the fit as noise at unit scale. This adds a std_floor parameter to fit_sparse_probe and sweep_sparse_probe (default 1e-3, matching the reference): each non-constant column's scale is now max(std, std_floor).

• Zero-variance columns keep scale one, as the guide already documents. Their coefficient is always zero, so the scale never reaches a prediction; scale one just avoids 0/0 when std_floor=0. This is the one remaining divergence from the reference, and it's noted in the guide's Reference section.
• std_floor=0 disables the floor and reproduces the previous behavior.
• std_floor is validated (finite, non-negative, not bool) and recorded on SparseProbeResult.
• The max_iter docstrings and the guide now state the LBFGS evaluation budget, max_iter * 5 // 4 (equal to max_iter when max_iter <= 3); the line search can go one evaluation past it.

Note: with the new default, preprocess="standardize" results change for columns whose training std is below 1e-3.

make unit-test and uv run mypy . pass locally. Happy to change the constant-column handling if you'd prefer it to follow the reference exactly.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• This change requires a documentation update

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Welcome to working on TransformerLens @VictorAraujopy! Thanks for putting this solution together. A couple small feedback items below before we get this merged

[VictorAraujopy]

Thanks for the review @jlarson4 , all three addressed in e1e13e3 ;)
The sweep test records the std_floor reaching every _selected_data call, so it also covers the control fits.

[jlarson4]

Thank you for resolving those for me @VictorAraujopy! Looks great, approved and merging now.