fix: optimize concurrency settings and update documentation for performance improvements

[ronibhakta1]

This pull request makes several improvements to CPU resource management, model selection, and logging for the PocketTTS-based synthesizer. The most significant changes are: defaulting to smaller, faster 6-layer models for most languages (except French), improving hardware-aware concurrency configuration, clarifying documentation around model sizes and CPU usage, and enhancing logging with synthesis timing details. These updates ensure better performance, more predictable resource usage, and clearer configuration/setup for users.

Concurrency and resource management improvements:

• The default value for max_concurrent_syntheses is reduced from 2 to 1 to avoid CPU oversubscription on small hardware; detailed comments clarify that each synthesis uses ~2 cores, and configure.sh now derives optimal values from detected hardware. [1] [2] [3] [4]
• Documentation (configuration.md) is updated to explain that each synthesis job uses ~2 CPU cores, and that concurrency is now hardware-derived rather than a static default.

Model selection and documentation:

• The language-to-model mapping in _LANG_MODEL now defaults to the 6-layer model for all supported languages except French, which only has a 24-layer model; this is explained in both code comments and documentation. [1] [2] [3] [4] [5]
• Documentation is clarified to show which model variant is used for each language, including a new table in the provider docs.

Performance and logging enhancements:

• Synthesis logging is improved to include start/finish messages and timing for generation and encoding, aiding performance monitoring and debugging. [1] [2] [3]

Threading and provider loading:

• The provider loader no longer sets torch.set_num_threads(1) because PocketTTS already manages this internally to optimize CPU usage.

These changes together make the system more efficient, easier to configure for different hardware, and clearer for users and developers.