Summary

This PR increases FSG phonetic context support from 128 to 256 phones and simplifies the bitvector implementation for better maintainability. The motivation is that models with large numbers of phones otherwise will fail -- in particular, Chinese -- without recompilation.

Changes

• Increase FSG_PNODE_CTXT_BVSZ from 4 to 8, supporting up to 256 phones (up from 128)
• Replace compile-time conditional macros with generic implementation
• Improve error message to show actual vs. maximum phone count

Rationale

The previous implementation required recompilation to support larger phonesets. This change:

1. Supports larger phonesets: 256 phones covers all known phonetic inventories without custom builds
2. Simplifies maintenance: Removes 20+ lines of conditional macro code in favor of generic function
3. Negligible performance impact: Bitvector operations occur during word transitions, not in the acoustic scoring hot path (~0.002% overhead)

Memory Impact

• Per fsg_pnode_t: +16 bytes (11% increase from ~144 to ~160 bytes)
• Typical grammars: +160KB to 1.6MB for 10K-100K nodes
• Acceptable tradeoff for runtime flexibility

Testing

• Clean compilation with no errors
• Existing unit tests pass
• Backward compatible with existing models and grammars