Rewrite GNU3 demangler for performance using DemangledTypeNode

Replace the TypeBuilder-based demangling path with a lightweight
DemangledTypeNode representation that defers type object construction
until the symbol is fully parsed. This avoids repeated heap allocation
and ref-count churn during recursive descent.

Key changes:
- Add DemangledTypeNode / demangled_type_node.{h,cpp}: a compact IR
  that mirrors the type grammar without allocating BN Type objects
- Use a thread_local demangler instance to amortize vector allocations
  across calls
- Also commonize some of the demangled string length calculations.

Result: ~3x throughput improvement on a 180K-symbol corpus with
97.7% success rate (matching the previous implementation).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: plafosse