fix memory leak: ensure hooks are cleaned up on exception in find_alignment

[virajsabhaya23]

Fixed Critical Memory Leak in Word-Level Timestamp Alignment

Problem Identified

• Memory leak in find_alignment() function causing hooks to persist on exception
• PyTorch forward hooks not cleaned up when errors occur during inference
• Accumulated hooks degrading model performance over time

Root Cause Analysis

• Location: whisper/timing.py, lines 187-204
• Issue: Hook cleanup code placed outside exception handling
• Impact:
  ○ Hooks remain attached to cross-attention layers on failure
  ○ Memory consumption increases with each failed alignment attempt
  ○ Model state becomes polluted across subsequent inference calls
  ○ Performance degradation in production environments

Solution Implemented

• Wrapped hook registration and forward pass in try/finally block
• Before: Hooks cleaned up only on successful execution
• After: Hooks guaranteed to be removed regardless of exceptions
• Ensures proper resource cleanup in all code paths

Technical Details

• Modified function: find_alignment()
• Protection added:
  ○ try block wraps model forward pass and attention weight extraction
  ○ finally block ensures all registered hooks are removed
  ○ Exception-safe cleanup prevents memory leaks
  ○ No performance impact on normal execution path