Restructure metrics reference documentation into category-specific subpages

[gaga1313]

This PR restructures the flat evaluation metrics page in the Metrax documentation. Instead of listing all metrics on a single, long page, they are now organized into dedicated, logical subpages corresponding to their application domains.

The documentation generation process is now automated using a newly created hook in conf.py which is run during sphinx-build.

• docs/base.rst: General helper metrics (e.g., Average).
• docs/classification.rst: Binary and multi-class evaluation metrics (e.g., Accuracy, Precision, Recall, AUCPR, AUCROC, FBetaScore).
• docs/regression.rst: Regression metrics (e.g., MAE, MSE, RMSE, RSQUARED).
• docs/ranking.rst: Information retrieval and search ranking metrics (e.g., AveragePrecisionAtK, DCGAtK, NDCGAtK, MRR, PrecisionAtK, RecallAtK).
• docs/nlp.rst: Natural Language Processing metrics (e.g., BLEU, Perplexity, RougeL, RougeN, WER).
• docs/image.rst: Computer vision metrics (e.g., Dice, IoU, PSNR, SSIM).
• docs/audio.rst: Audio metrics (e.g., SNR)

Additionally, several public metrics exported in the metrax namespace (CosineSimilarity, MSLE, RMSLE, SpearmanRankCorrelation) but previously undocumented have now been added to their respective categories.