Add ReadOnlyMemory<char> support

[danipen]

Summary

This PR introduces ReadOnlyMemory<char> support throughout the tokenization pipeline, enabling zero-allocation text handling. Combined with several allocation reduction optimizations, this delivers significant performance and memory improvements.

What's New

New LineText Type

A new LineText struct wraps ReadOnlyMemory<char>, providing a clean API for text handling without string allocations:

// Before: Always allocated strings
string lineText = model.GetLineText(lineIndex);
grammar.TokenizeLine(lineText, ruleStack, timeout);

// After: Zero-copy memory access
LineText lineText = model.GetLineText(lineIndex);
grammar.TokenizeLine(lineText, ruleStack, timeout);

Updated Public APIs

• IGrammar.TokenizeLine() and TokenizeLine2() now accept LineText instead of string
• IModelLines.GetLineText() returns LineText instead of string
• Implicit conversions from string to LineText maintain backward compatibility

Benchmark Results

Tested with a 133,439 line C# file (5.8 MB):

Metric	master	This PR	Improvement
Execution Time	4.752 s	2.681 s	44% faster
Memory Allocated	658.27 MB	496.36 MB	25% less
Gen0 Collections	82,000	62,000	24% fewer
Gen1 Collections	8,000	4,000	50% fewer

Optimizations Applied

1. ArrayPool for line buffers - Reuse char arrays instead of allocating per line
2. Allocation-free timing - Use Stopwatch.GetTimestamp() instead of new Stopwatch()
3. List pooling - Reuse internal lists in hot paths (HandleCaptures, CheckWhileConditions)
4. Scope name caching - Cache GetScopeNames() result to avoid repeated list creation
5. Single-scope optimization - Avoid List<string> allocation for single scope pushes

⚠️ Breaking Changes

• IModelLines.GetLineText() now returns LineText instead of string
• IGrammar.TokenizeLine() signature changed to accept LineText

Migration: Replace string with LineText in implementations. The implicit conversion from string means most call sites work unchanged.

Testing

• New benchmark project for performance validation