fix: harden cache init and nlp metrics edge cases

[AhmedAli58]

Summary

This PR resolves three meaningful quality issues focused on correctness and reliability:

1. Import-time cache path failure in restricted environments

• pyhealth could fail to import when ~/.cache/pyhealth was not writable.
• Added resilient cache initialization with preferred/fallback paths, better error context, and PYHEALTH_BASE_CACHE override support.
• Made logger handler setup idempotent to avoid duplicate handlers.

2. Broken missing-dependency handling in NLP scorer

• Scorer._get_missing_modules() used variable shadowing and invalid accumulation on install-failure paths.
• Reworked missing-module collection so unavailable methods are removed cleanly and missing requirements are returned consistently.

3. Empty score serialization crash

• ScoreSet.as_numpy() raised on empty results due to np.stack([]).
• Added safe empty-array handling so empty score sets serialize to valid numpy/dataframe outputs.

Files Changed

• pyhealth/__init__.py
• pyhealth/nlp/metrics.py
• tests/core/test_cache_path.py
• tests/nlp/test_metrics_edge_cases.py

Validation

• PYHEALTH_BASE_CACHE=/tmp/pyhealth-cache pytest -q tests/core/test_cache_path.py tests/nlp/test_metrics_edge_cases.py
• python -m py_compile pyhealth/__init__.py pyhealth/nlp/metrics.py tests/core/test_cache_path.py tests/nlp/test_metrics_edge_cases.py

Impact

• Improves robustness of package import behavior across constrained environments.
• Strengthens dependency-failure paths in NLP scoring utilities.
• Prevents crashes for valid empty-input workflows and adds regression coverage for all three fixes.