Add analyze_weights API for per-layer weight-entry distribution fitting

[charlesmartin14]

Motivation

• Provide a dedicated, testable routine to analyze the distribution of raw weight entries per layer and compare left/right tails to multiple candidate families (power law, exponential, log-normal, Laplace).
• Keep the new analysis isolated from the main analyze() workflow by placing it in a separate module so it can be unit tested in isolation.
• Support practical requirements like fast subsampling and optional plotting for interactive inspection of fits.

Description

• Added a new module weightwatcher/analyze_weights.py that implements _ks_metrics, _fit_side_models, _plot_side_distribution, and analyze_weights to fit left/right sides of the flattened weight entries to power_law, exponential, lognormal, and laplace using scipy.stats.
• Fit scoring combines the Kolmogorov–Smirnov statistic and log-likelihood, and the code marks a single best fit per side with is_best_fit and best_fit_for_side fields; results are returned as a pandas.DataFrame and stored in watcher.details.
• Added a wrapper method WeightWatcher.analyze_weights(...) in weightwatcher/weightwatcher.py that follows the analyze()-style signature and delegates to the new module so callers can invoke the new analysis via the WeightWatcher object.
• Added unit tests tests/test_analyze_weights.py which generate synthetic samples for each target family (power-law, exponential, log-normal, Laplace) and assert that _fit_side_models selects the expected best distribution.

Testing

• Ran the unit tests for the new analyzer with pytest -q tests/test_analyze_weights.py.
• Result: 4 passed (all synthetic-distribution tests succeeded).

---

Codex Task