Add structured logging and feature output diagnostics to FeatureBuilder

[sundy1994]

Summary

• Introduces a reusable setup_logger utility in src/team_comm_tools/utils/preprocess.py that writes timestamped logs to ./<output_file_base>/logs/, auto-creates the directory, and guards against duplicate handlers / propagation to root.
• Wires two loggers into FeatureBuilder — feature_builder.log for top-level run info and summary_details.log for verbose per-column output — and threads the logger through ChatLevelFeaturesCalculator, UserLevelFeaturesCalculator, ConversationLevelFeaturesCalculator, and check_embeddings. print and warnings.warn calls for errors / invalid configs are now mirrored to the log.
• Adds perf_counter-based timings around each feature method (chat / user / conversation level) and around sentence-vector and BERT generation in check_embeddings.py, so the log captures per-step durations.
• Adds a post-featurization diagnostics step (generate_summary_stats) that, for each output level, reports columns with high NA ratios, high zero ratios, and groups of highly correlated columns (Spearman, configurable threshold). New constructor params: corr_thresh, min_na_ratio, min_zero_ratio, min_group_size, treat_zero_as_na, drop_redundant_columns. With drop_redundant_columns=True, columns with NAs/zeros that exceeding the thresholds are dropped. Moreover, only one representative per correlated group is kept (chosen by valid-data count and variance) and others in the group are dropped.
• Logs the run header (timestamp, dataset shape — lines / unique speakers / unique conversations) at the start of featurize.
• Cleans up imports in check_embeddings.py (drops top-level torch / unused util in favor of narrower from torch import cuda, no_grad).
• Updates and reorganizes docstrings in feature_builder.py.
• Adds *.csv and *.log to .gitignore.

Behavior change to flag for review

As discussed, the first x percent feature is deprecated. analyze_first_pct and get_first_pct_of_chat are commented out in feature_builder.py, removing the multi-truncation loop in featurize.